Blue-Collar AI: Wearable Camera + Real-Time Guidance

Market assessment for Wenhao Dong & Eric San · B2B vertical SaaS with hardware
11 February 2026 · R3: Construction + Sweep Update

Wenhao wants to build a wearable camera platform for blue-collar field workers — AI-powered quality control and real-time guidance for hotel housekeeping, aircon repair, and maintenance crews. Eric would build the MVP part-time. This assessment covers market size, competitive landscape, unit economics, and a founder-contextualized GTM.

Field Service Mgmt
US$6.2B
11% CAGR1
Connected Worker
US$5.5B
→ US$18B by 20302
AI Visual Inspect.
US$1.8B
28% CAGR3
YC RFS Match
✓
"AI for Physical Work"5

I. Who We Are (for context)

Wenhao DongEric San
Sales-first founder Based in Singapore. Currently running Smilie — a corporate gifting startup. Very operational, growing gradually, but hitting a ceiling: pivoting to software for scale/margin is proving extremely difficult, and regional contracts are blocked because gifting is localized (local office approves local spend, not regional). Previously owned a printing company which he sold. Has warm connections in aircon/repair, home maintenance, hotel management, printing. Knows factory and ops pain firsthand. "Not gonna do this as a side hobby shit, i see this as my next phase." Technical co-founder (part-time) Based in Hong Kong. Uber (onboarding drivers), then Pickupp (onboarding couriers, walkers, riders across HK, SG, KL, HCMC — started in operations handling the whole ops function, then pivoted to tech/products). Now building GenieFriends (event ops, decorations, restaurant partner management) and Donna AI. Hardware tinkering (Blackring). Bandwidth: intention is to settle on one good project — currently exploring/validating to commit yes or no, instead of hanging a project halfway. ~10 hrs/week initially.
Founder dynamic: Wenhao is the operator/seller. Eric is the builder. This is the right pairing for B2B hardware+software. Wenhao's Smilie experience is instructive: he's living the pain of an operational business that can't scale past local, which is why software-first matters for his next venture.
Eric has literally built this before. At Pickupp, the core operational challenge was: how do you verify that freelance delivery agents are doing their job correctly? The solution Eric helped build:
  • Agents take photos at pickup + delivery to prove they handled the right package
  • Agents scan QR codes to verify package identity
  • GPS tracking to confirm on-time pickup, on-time dropoff, correct location
  • All feeding into a dashboard for ops managers to monitor compliance
This is exactly the same architecture Wenhao's venture needs: field worker → photo/scan → verification → manager dashboard. The only addition is AI vision replacing manual review. Eric isn't learning a new pattern — he's upgrading one he already built across 4 markets (HK, SG, KL, HCMC). This is a 0→1 insight, not a guess.
The commitment question (from Eric's own notes26):
Eric is exploring multiple avenues — Donna AI, Blackring, avet, GenieFriends wind-down. He's explicitly stated: "Not advisor role — real validation." And: "Wants to validate + protect downside."26 His concern: "How to land early revenue? Eric hesitates on B2B that doesn't earn."26

This is honest and healthy. The repair.sg S$10K angel investment doesn't solve runway, but it signals conviction — Zames is putting money and workers on the line. The real proof comes from early revenue. The deal: Eric validates the MVP at 10 hrs/week through Phase 2. If it works, he commits or they hire. If it doesn't, clean exit. No project left hanging.

II. The Thesis

Two products built on a wearable camera platform for blue-collar field workers:

ProductWhat It DoesDifficultyValue Prop
Product 1: Prevention Real-time visibility of field operations. Camera on worker → monitor, detect anomalies, prevent theft/fraud. Easier to build Harder to sell — "security camera is good enough" until the US$10K hotel room theft happens
Product 2: Support Buddy AI-powered real-time QC + guidance. "You forgot a screw." "Bedsheet corner left untucked." Technical support without a second person. Harder to build — requires multimodal AI + domain-specific training Easier to sell — directly replaces supervisor time, measurable ROI
Wenhao's instinct is right: Product 2 (Support Buddy) is the real business. Product 1 is a feature, not a product — the security camera market is commoditized. Product 2 is what YC is signaling as the "alpha."5

Which vertical first? TO VALIDATE

VerticalSG Market SizeTask RepeatabilityVisual QC FitWenhao's AccessWTP Signal
Gift/package packing QC Tiny — ~200–500 packers across all SG corporate gifting ops. US$48–120K ARR at 100% pen. Very high — same items, same arrangement, same wrapping per SKU Strong — right items? right layout? wrapping intact? IS the customer — Smilie packers Firsthand pain — Wenhao has this problem TODAYTW5
Hotel housekeeping Small — ~3,000–4,500 workers. 281 gazetted hotels.22 ~US$720K–1.1M ARR at 100%. Very high — same 20+ checkpoints every room Strong — visual standards (bed, bathroom, amenities) Friends in hotel mgmt Unknown — need to validate
Commercial cleaning Large — ~85,000 workers.21 NEA-regulated. US$20M ARR at 100%. High — standardized per contract (office, retail, hospital) Strong — floor, surfaces, restrooms = visual checkpoints Adjacent to hotel contacts Unknown
F&B kitchen / restaurant prep Very large — ~150,000 workers.21 SFA food safety compliance. ~US$36M ARR at 100%. High — plating, portion, cleanliness, food safety Strong — visual standards are SFA-regulated Eric's GenieFriends restaurant partners Crowded — Agot.AI16, Wobot.AI17 already in market.
Logistics / warehouse Medium — ~50,000 workers.21 SG is a regional hub. ~US$12M ARR at 100%. High — pick/pack/scan/sort are highly repetitive Strong — visual verification of picks, packing, labels No direct access Unknown — subsegments below
Aircon / HVAC / general repair Small — ~5,000–8,000 technicians in SG.21 Medium — varied jobs, less standardizable Medium — some visual, but also electrical/mechanical Direct friends + Zames Chew (repair.sg)25 Validated — Zames: S$40–50/worker/mo acceptable25
Printing QC Niche — ~2,000–3,000 print workers in SG. Declining sector. High — color matching, alignment, defect detection Strong — visual inspection is the entire QC process Wenhao's old industry Unknown — but Cognex/Keyence proves the market13
Construction site safety / compliance Very large — ~300K+ workers (CMP work permits: 456,800 as of Dec 202427; construction is ~60–70% of CMP). Largest vertical in this table. US$72–84M ARR at 100%. High — PPE checks, scaffold inspections, housekeeping, hot-work permits are standardized across all sites Strong — safety hazards are highly visual (no helmet, missing barricade, unsecured scaffolding, standing water) 2 direct contacts — one in building construction (~100 workers), one in building maintenance (~100 workers)28 Strong verbal, needs validation — Contact described product as "blue-collar compliance assistant on the field."28 No pricing conversation yet.
Honest vertical sizing: Gift packing QC is a tiny market — at best a few hundred packers in all of SG. It's an excellent dog-food/MVP opportunity, but it's not venture-scale. The volume verticals are construction (300K+ workers), F&B (150K workers), commercial cleaning (85K workers), and logistics (50K workers).2127 Hotels are a strong-signal niche but small. The MVP vertical and the scale vertical may be different — start on Smilie to validate AI + workflow, but don't plan around gift packing as the business.

Feb 11 update: Construction emerges as the largest addressable vertical in this table by workforce size. It also has the strongest compliance driver (WSH Act29) and the richest government grant stack (BETC up to 70% for SMEs31). However, construction sales cycle is longer (main contractor is the buyer, not individual companies)36 — see GTM assessment below. No decision to commit to construction yet — it needs the same WTP validation as the other verticals.
Zames Chew (repair.sg) — only externally validated WTP in this entire report25
At the Song Fa dinner (Feb 1), Wenhao tested the pitch with Zames Chew (repair.sg founder). His feedback:
  • WTP: S$40–50/worker/month — "if useful, cost is justified." This is ~US$30–37, higher than our baseline US$15–25 estimate.
  • Would deploy to his full-time workers — this is a pulled product, not pushed.
  • Agreed to invest S$10K as angel (per Eric — not yet documented in formal notes). Skin in the game = genuine conviction.
  • Privacy concern raised: clients may not allow cameras at their premises. This is a blocker for on-site repair work specifically.
  • Wenhao believes repair.sg is a better first pilot than Smilie.
Critical assessment: This is the only external WTP signal in the report, and it's strong. But we should stress-test it:
  • One dinner conversation ≠ purchase order. The founder said "if useful" — conditional on seeing it work. Verbal WTP at a friendly dinner is 3–5× inflated vs actual buying behavior.24
  • The privacy blocker is real. Repair workers go to clients' homes and offices. If clients object to cameras, the product literally cannot be used. This is a different dynamic from packing QC (Wenhao's own warehouse) or hotel housekeeping (hotel's own rooms).
  • Workflow variation is real but solvable. Repair jobs vary (aircon ≠ plumbing ≠ electrical) — but in 2026, each job type is just a different text prompt to the vision model, not a custom-trained model. "Check that the refrigerant line is connected" vs "check that the wiring is secured" is a checklist swap, not a re-engineering.
  • The angel investment signals conviction. S$10K is a modest check, but from a friend-founder who would also deploy his workers as a pilot. An investor-customer is an extremely strong Phase 1 signal, even if the vertical itself isn't the scalable one.
Construction site safety — Wenhao's emerging thesis (Feb 9)28
Wenhao: "I got a feeling we might do the construction... We see exactly what the construction workers are seeing. It's easier to understand what objects we need to map out." His construction contact framed the product as "a blue-collar compliance assistant on the field."

Why construction is compelling:
  • Compliance-driven buyer. WSH Act penalties: up to S$500,000 fine + S$20,000/day for stop-work order non-compliance, plus up to 12 months imprisonment.29 Safety isn't discretionary — it's legal obligation.
  • Quantifiable ROI. Wenhao's contact described 3–5 day site shutdowns after serious incidents. On a S$50M project, even one day of delay costs S$15–30K+ in prolongation costs (equipment idle, labor, overhead). A single prevented shutdown easily covers a year of SaaS.
  • Regulatory tailwind (June 2024). MOM now mandates video surveillance systems for construction sites with contracts ≥S$5M.30 The government is already pushing cameras onto construction sites.
  • Richest grant stack. BETC Grant: up to 70% for SMEs (Apr 2025–Mar 2027).31 PSG: up to 50%.23 WSH Council technology grants cover AI, wearables, and sensors.32
  • No privacy concern. Unlike repair (client homes) or hotels (guest rooms), construction sites are employer-controlled. Camera resistance is lower.
  • 2 warm contacts. One in building construction (~100 workers), one in building maintenance (~100 workers). Instant pilot pool.
  • 20 construction deaths in 2024 (47% of all SG workplace fatalities), plus 76 deaths + major injuries in H1 2025.33 Safety is a live, visceral problem.
Why construction is harder than it looks:
  • Main contractor is the buyer, not subcontractor. Safety tech decisions are centralized — the main contractor deploys systems that subcontractors must use.36 This means longer sales cycles and larger deal sizes than hotel/cleaning.
  • Worker transience. Different crews rotate through a project every few weeks/months. Training and adoption resets with each rotation.
  • Harsh environment. Dusty, wet, outdoor conditions. Phone/wearable durability is a real question for outdoor construction vs indoor packing/hotel.
  • Tech adoption barriers. 73% of SG firms cite high costs as the #1 barrier to technology adoption (SBF survey, 2024).37 44–47% cite skill gaps. Grants help but don't eliminate resistance.
  • Protex AI (YC S21) exists. US$36M Series B (Jan 2025), DHL/UPS clients, computer vision for safety on existing CCTV.34 Different approach (fixed cameras vs wearable) but addresses overlapping pain points. See competitive landscape.
No decision to commit to construction yet. It needs the same WTP validation as the other verticals. But the combination of market size (300K+ workers27), compliance driver (WSH Act29), and grant support (BETC 70%31) makes it worth serious investigation alongside cleaning and hotels in Phase 1b.

Logistics subsegments

Sub-verticalWhat Camera VerifiesExisting Camera Infra?Opportunity
Pick & pack (warehouse) Correct item picked, correct qty, correct box Partial — barcode scanners exist, but no visual verification of packed contents Strong — Smilie packing QC is literally this subsegment. Transferable.
Conveyor belt / sorting Correct routing, label readable, damage detection Heavy — Cognex/Keyence cameras already on most high-throughput belts13 Hard — incumbents (Cognex US$800M rev, Keyence US$7B rev) own this. Don't compete here.
Last-mile delivery proof Package delivered to correct location, photo proof, condition on delivery Yes — every delivery app already asks drivers to take photos. Eric built this at Pickupp. Saturated — feature of every delivery platform, not a standalone product.
Loading dock / receiving Shipment matches PO, damage on receipt, count verification Minimal — mostly manual clipboard-based today Medium — real pain but niche. 3PLs and distributors.

Key insight: "pick & pack" and "loading dock" are the addressable subsegments. Conveyor belt is owned by Cognex/Keyence. Last-mile delivery proof is a commoditized feature, not a product.

The Camera Startup Graveyard — Why F&B / Retail Camera AI Is Treacherous

Camera-based ops analytics was a hot category in 2019–2023. The results were mixed at best:

CompanyWhat They BuiltOutcomeLesson
Presto (Palo Alto)20 Restaurant automation — drive-thru AI, table-side ordering. SPAC'd as PRST. Struggled — Revenue missed targets, stock cratered post-SPAC. AI accuracy in noisy kitchen/drive-thru environments was insufficient. Restaurant AI is harder than it looks. Noisy, varied environments break generic models.
Dragontail Systems19 Kitchen management AI — overhead camera watches food prep + packaging. Detects order completion. Acquired by Yum! Brands subsidiary19 — tech was good enough for acquisition but couldn't build standalone business. Kitchen camera QC works technically but needs a distribution channel (franchise parent). Standalone is hard.
Agot.AI16 Kitchen QC camera for QSR — overhead camera monitors line, detects wrong orders, missing items. Chick-fil-A, others. Raised ~US$11M Series A (2023). Still active.16 The survivor. Used EXISTING overhead cameras, targeted enterprise QSR chains, deep single-vertical focus. Donna R3: no verified expansion into hospitality or cleaning.TW10
Wobot.AI17 Video intelligence for restaurant + retail + manufacturing ops. AI layer on existing CCTV cameras. Raised ~US$10M. Active (India/US).17 Uses EXISTING security cameras — no new hardware. Multi-vertical platform. ~US$50–100/camera/mo. Donna R3: expanding to manufacturing/warehousing, not hospitality.TW10
Verkada18 Connected security cameras with analytics. US$3.2B valuation.18 Dominant in connected cameras. But security-first, ops QC is secondary feature. If Verkada adds kitchen/cleaning QC analytics, they own the camera AND the AI. Hard to compete.
The honest pitch challenge for F&B: A restaurant considering Wenhao's tool will ask: "We already have security cameras + Wobot/Verkada analytics. Why add another camera system?"

Pattern from survivors: Agot.AI won by (a) using EXISTING cameras, (b) going deep on ONE vertical (QSR kitchen), (c) selling to enterprise chains, not SMBs. Wobot.AI won by (a) being a software layer on ANY existing camera, (b) multi-vertical but light-touch analytics, not deep QC.

What this means for Wenhao: F&B kitchen is treacherous as a first vertical. Camera infrastructure already exists (CCTV), and funded players (Agot, Wobot) are ahead. Packing QC and hotel housekeeping are better first verticals because (a) there are NO existing cameras pointed at packing stations or hotel rooms, and (b) no funded competitor is focused on these specific QC tasks. Donna R3 confirms: Agot locked in QSR, Wobot expanding to manufacturing — neither targeting hospitality/cleaning QC.TW10
Existing camera infrastructure by vertical:
  • Retail / F&B: Heavy — security CCTV everywhere. Verkada18, Wobot17 already adding AI analytics. Hard to wedge in.
  • Logistics (conveyor belt): Heavy — Cognex/Keyence13 industrial vision cameras. US$2K–10K per unit. Dominant.
  • Logistics (pick & pack station): Minimal — barcode scanners but no visual verification camera. Gap exists.
  • Hotel rooms: None — no cameras in guest rooms (privacy). Worker brings the camera. This is the gap.
  • Cleaning sites: None — no cameras in offices/hospitals being cleaned. Same gap.
  • Packing stations (SMB): None — Smilie-size operations don't have industrial vision cameras. Gap is wide open.
  • Construction sites: Partial — MOM mandates CCTV for sites ≥S$5M (June 202430), typically at entrances/perimeters. But zero cameras tracking individual worker behavior at height, in active zones, or on scaffolding. Protex AI34 uses existing CCTV for facility-level safety detection — but nobody is doing worker-level wearable/phone vision on construction sites in SEA. The gap is worker-level visibility, not site-level surveillance.
Rule of thumb: go where there are NO existing cameras. That's where you're adding infrastructure, not competing with it. Construction sites have some CCTV but a massive gap at the worker level — the mandatory site cameras don't see what the worker at level 4 sees.
Don't pick a winner yet — validate in parallel.
Smilie packing is the easiest place to start (Wenhao IS the customer, zero sales cycle), but it's a tiny market. The real business likely lives in one of the volume verticals: commercial cleaning (85K workers21), F&B (150K21), or logistics (50K21). Hotels are a strong-signal niche but small (3–4.5K workers22).

Recommended validation plan: (1) Dog-food on Smilie packers to validate the AI + workflow work at all. (2) Simultaneously run WTP conversations across hotel, cleaning, and logistics contacts. (3) Let the data decide. Don't commit to a vertical until you have signal from at least 2–3 real conversations per vertical. The worst outcome is building for a market too small to sustain a venture.5

III. Market Sizing (Layered)

LayerMarketSize (2024)Projected (2030)CAGR
Global TAM Field Service Management Software US$6.2B US$11.5B ~11% 1
Segment TAM Connected Worker Platforms US$5.5B US$18B ~18% 2
Segment TAM AI Visual Inspection / QC US$1.8B US$8B ~28% 3
Regional TAM SEA Facilities Management Tech ~US$500–800M ~US$1.5–2B ~15–18% 4 (est.)
Addressable TAM SMB field service in SG + APAC (repair, hotel, logistics) ~US$50–150M ~US$200–500M Calculated
Field Service Mgmt
US$6.2B
Connected Worker
US$5.5B
AI Visual Inspect.
US$1.8B
SEA Facilities Tech
~US$650M
Our addressable
~US$100M
Key insight: "Connected Worker Platform" (US$5.5B → US$18B)2 is the actual category Wenhao is building in. This is not a consumer play — it's B2B SaaS + hardware sold to operations managers. The 18% CAGR is one of the fastest in enterprise software.2 YC validated5

Google Trends signal (Donna R2): "Visual Inspection AI" queries are outscaling "Quality Control Software" in SEA by 3.2× as of early 2026.TW9 The market is searching for AI-specific solutions, not just generic QC tooling. This is a category-creation signal.

Singapore-Specific Sizing (GTM-First Lens)

Global TAM is interesting but the GTM starts in SG. What's actually addressable from where Wenhao sits?

SectorSG Workforce (est.)Visual QC FitAt US$20/worker/moNotes
Hotels (housekeeping) ~3,000–4,50022 Strong US$60–90K/mo at 100% pen 281 gazetted hotels (STB 2024)22, ~67K rooms, ~1 housekeeper per 15–20 rooms. Donna R2: @rentokofficial deploying real-time ops for student housing — validates shift from manual to digital.TW6
Commercial cleaning / FM ~85,00021 Strong US$1.7M/mo at 100% pen NEA-regulated. Office, retail, hospital cleaning. 20× larger than hotels.21
F&B / food service ~150,00021 Medium-Strong US$3M/mo at 100% pen SFA-regulated food safety.21 Kitchen prep, plating, cleanliness. High compliance cost.
Logistics / warehouse ~50,00021 Strong US$1M/mo at 100% pen SG is a regional logistics hub.21 Pick/pack/scan verification.
Total SG addressable ~289,000 workers US$5.8M/mo at 100% At 1% penetration = ~US$700K ARR. At 5% = ~US$3.5M ARR.
The SG-to-SEA expansion math: SG is the beachhead but SEA is the TAM multiplier. Malaysia alone has ~2× the hotel workforce, Thailand ~3×, Indonesia ~10×. A product that works for SG hotel housekeeping works for Kuala Lumpur, Bangkok, Jakarta with localization. At 1% SEA penetration across hotels + cleaning + F&B, the addressable ARR is US$10–30M. This is where the venture becomes fundable at Series A scale.

But GTM depends on the vertical: If packing QC is the first product (Smilie → other SMB fulfilment), the TAM is smaller but the sales cycle is shorter (Wenhao sells to people like himself). If hotel housekeeping is first, the TAM is larger but requires industry-specific sales. The vertical choice changes the market sizing.
SG Government Grant Lever: PSG (Productivity Solutions Grant) covers up to 50% of qualifying digital solutions, capped at S$30K per solution.23 EDG (Enterprise Development Grant) covers up to 50% for business upgrades.23 If Wenhao's tool gets PSG pre-approved, the effective price to the buyer halves — a US$20/worker/month tool effectively becomes US$10/worker/month. This is a massive GTM accelerant in SG. Getting on the PSG pre-approved list should be a Phase 2 goal.

Caveat: PSG pre-approval requires a working product + registered SG company. Timeline: ~3–6 months to get approved.

Donna R3 confirmed: The IMDA PSG pre-approved list already includes CleanSmart, WhizWork, Infogrid (cleaning management) and Novade, Kegmil (field service).TW10 However, these are scheduling/IoT-based tools — no Vision-based QC layer exists in the current PSG list. This validates AI visual QC as a differentiated entry point that wouldn't compete directly with pre-approved incumbents but could ride the same grant funding channel.

IV. Competitive Landscape — Who's Already Here

4a. Enterprise Players (US$50M+ raised)

CompanyHQRaisedModelWhat They Got RightWhy They're Not Us
Augmentir6 Austin, TX US$45M+ AI-powered connected worker. Digital work instructions + skills tracking. Feb 2026 update: 38% team growth in H2 2024, new clients include Colgate-Palmolive, Mondelēz, Duracell, Hitachi Energy. Named leader by Frost & Sullivan (Jan 2026). 5M+ AI-optimized time/motion studies. Added AR features (Sep 2025) and video-to-procedure GenAI conversion.6b Category leader for manufacturing connected worker. Expanding fast with GenAI features. Different vertical — focused on manufacturing (discrete, process). Not maintenance/field service/construction. Zero Twitter/Reddit mindshare despite being category leader — suggests enterprise-only distribution, not PLG.
Parsable (→ GE Digital)10 San Francisco US$80M+ Connected worker for industrial ops. Acquired by GE Digital 2023. Proved enterprise buyers pay US$20–50/worker/month for connected worker platforms Got acqui-hired into GE ecosystem. Validates exit path for smaller players.
RealWear8 Vancouver, WA US$180M+ Industrial smart glasses (Navigator Z1). Hardware + software. Proved hardware-first connected worker model works. US$3,000+ per headset. Hardware trap — US$3K/device limits adoption to high-value industrial. Wenhao's US$20 camera is 150× cheaper.
Tulip12 Boston US$100M+ Composable MES. Frontline operations platform. No-code approach lets operations teams build their own apps Manufacturing-centric. No camera/vision component. Different buyer.
SightCall11 Paris / SF US$55M+ Visual assistance for field service. AR overlays via phone/tablet. Uses existing phone cameras — no custom hardware needed Closest competitor to Product 2 thesis — but requires manual initiation (worker calls for help), not always-on AI.
Protex AI34 Limerick, IE US$36M (Series B, Jan 2025) Computer vision on existing CCTV for workplace safety. Detects PPE violations, forklift speed, walkway adherence, 100+ hazard types. DHL achieved 64% risk reduction in 3 months. Clients: DHL, UPS, FedEx, GEODIS, M&S. YC S21. Deloitte Fast 50 Rising Star (Dec 2025).34 Proved AI safety detection works and Fortune 500 will pay. Privacy-preserving approach. Strong traction. Overlapping pain point — addresses same safety/compliance need. But different product: Protex uses fixed CCTV (facility-level surveillance), not wearable/phone (worker-level guidance). Protex tells managers "this area is unsafe"; Wenhao tells workers "your harness isn't attached." Not in SEA. Not in construction (logistics/ports/manufacturing focus). Enterprise-only pricing.
SafetyCulture (iAuditor)35 Sydney, AU US$297M raised; ~US$2.7B valuation (Aug 2023) Mobile safety inspection checklist platform. 75K+ businesses, 1.5M+ workers, 180 countries. ~US$132M revenue (2022). Bottom-up PLG model: free tier → organic adoption → monetize on features. Started with construction/industrial inspections, expanded to all industries.35 Best GTM playbook analog. Won by being phone-first, free to start, zero-hardware, bottom-up adoption by safety officers. 85% ARR growth. Proves safety inspection is a massive market. Digital checklists, NOT AI vision. SafetyCulture is a form-filling tool — humans inspect and check boxes. Wenhao adds AI that sees what the human sees and flags issues automatically. Could be complementary (integrate with SafetyCulture checklists) or competitive (replace the manual checklist with AI vision).

4b. SEA / Regional Players

CompanyHQModelRelevance
Kegmil7 Singapore Field service SaaS. Checklists, work orders, maintenance scheduling. CMMS. ~US$20–30/user/mo.TW10 Feb 2026 update: Raised SGD$2.2M pre-Series A. Actively expanding into construction, manufacturing, marine, renewables.7b Direct overlap, growing — same market (SG field service), same buyers (facility managers), now also targeting construction. But still no AI, no camera, no real-time guidance. PSG pre-approved.TW10 Wenhao's AI vision layer leapfrogs this, but Kegmil has distribution advantage (PSG, existing customer base).
Novade15 Singapore Construction + facilities management. Smart forms, safety, quality. ~US$35–50/admin/moTW10 Closest SG competitor for construction vertical. Construction-focused. Enterprise pricing. PSG pre-approved.TW10 Validates SG buyers will pay for digital construction tools. But no AI, no vision, no cameras — digital forms only. If Wenhao targets construction, Novade is the incumbent to unseat or integrate with.

4c. Construction-Adjacent Players (Global)

These companies operate in construction but address different aspects than worker-level safety/compliance. Included because Wenhao may encounter them in construction buyer conversations.

CompanyHQRaised / ValuationWhat They DoRelevance to Wenhao
Buildots38 Tel Aviv US$166M total; US$300M val (May 2025) AI-powered construction progress tracking from 360-degree cameras. Clients: Turner, VINCI, Intel. Reduces project delays by up to 50%. Different use case — tracks construction completion (is the wall built?), not worker safety (is the worker wearing a harness?). Seven-figure enterprise deals, multi-year sales cycles. Not competing for the same budget.
Versatile (CraneView)39 Tel Aviv US$100M+ (US$80M Series B, 2021) AI device mounted under crane hooks. 12M+ crane picks analyzed. Monitors utilization, load, progress. Fixed infrastructure — crane-mounted, not wearable. Different form factor, different data (material flow vs worker behavior). Not competing.
Newmetrix (→ Oracle)40 Cambridge, MA → Oracle Acquired by Oracle (Oct 2022) AI safety detection from photos/video. "Vinnie" AI detects 100+ hazards. 1.2M safety tags generated for Obayashi. Now embedded in Oracle Construction Intelligence Cloud.40 Closest construction safety AI competitor globally. But acquired = locked inside Oracle ecosystem. Enterprise-only, US-focused. Pre-acquisition had only "a handful of sales reps." Couldn't scale standalone — validates that construction safety AI is hard to sell independently. Not accessible to SG SMBs on Oracle's cloud.
Procore Carpinteria, CA Public (PCOR). ~US$1.36B rev/yr (2025). 17K+ customers.41 Dominant construction management platform. Project management, financials, quality, safety modules. Integrates with Oracle/Newmetrix for AI safety. Platform, not competitor. Procore is the OS of construction; Wenhao's product would need to integrate with it, not fight it. Too large and enterprise to compete on safety specifically.
Competitive gap (updated Feb 11): No one in SEA is doing wearable camera + AI real-time guidance for field workers. This claim was confirmed across Reddit, Product Hunt, Twitter/X, and YC directory searches (Feb 2026) — zero organic discussion or products found for hotel housekeeping QC, warehouse packing verification, or cleaning AI across all four platforms.

The new wrinkle: Protex AI34 (YC S21, US$36M Series B, Jan 2025) does AI safety detection on existing CCTV for logistics/manufacturing. They address an overlapping pain point (workplace safety) but with a fundamentally different product: fixed cameras watching areas vs wearable cameras guiding workers. Protex tells managers "this zone is unsafe"; Wenhao tells workers "put your harness on." The differentiation is facility-level surveillance vs worker-level guidance. In a construction buyer conversation, Wenhao should expect to be asked about Protex. The answer: they complement each other — Protex monitors the site, Wenhao guides the individual worker.

Augmentir6 and SightCall11 remain US-focused, enterprise-priced (US$50+/worker/month). Kegmil (~US$20–30/user/moTW10) and Novade (~US$35–50/admin/moTW10) are the closest in geography but are traditional SaaS tools with no AI vision layer. Both are PSG pre-approved for scheduling/IoT — not for visual QC. Wenhao's US$15–25/worker/mo undercuts both while adding a capability they don't have.

V. The Graveyard — Who Tried and Failed

CompanyRaisedWhat They BuiltWhy They DiedLesson for Wenhao
Daqri (LA)9 US$275M Industrial AR smart helmet. Heads-up display for construction/manufacturing workers. Shut down 2019. Hardware-first strategy was too expensive (US$5K+ per helmet). Couldn't find repeatable enterprise deal flow. Burned cash on hardware R&D before proving product-market fit. CRITICAL — This is Wenhao's biggest risk. Hardware-first = high burn, slow iteration. Daqri spent US$275M and never found PMF.
Atheer (Mountain View) US$10M+ AR glasses for industrial workers. Remote expert guidance. Couldn't scale. Acquired by PTC for technology, not for the business. Acqui-hire exit — IP was valuable, business wasn't. PTC wanted the AR tech for Vuforia.
Scope AR (Edmonton) US$10M+ AR remote assistance for industrial workers. Pivoted multiple times. Couldn't differentiate from Zoom with camera. "Remote expert" is a feature, not a product. Only works if AI replaces the remote expert entirely.
The Daqri pattern: Every company that led with expensive custom hardware for industrial workers either died or got acqui-hired.9 The survivors (Augmentir6, Parsable10) were software-first — they worked on existing phones and tablets, then added hardware optionally. Wenhao's US$20 Alibaba camera is the right instinct — cheap, commodity hardware with AI value in the cloud. But the Daqri ghost still haunts: hardware supply chain risk is real.

VI. Unit Economics — Benchmarked

MetricWinner (Augmentir)6Regional (Kegmil/Novade)7TW10Our EstimateNotes
ARPUUS$30–50/worker/moKegmil: ~US$20–30/user/mo
Novade: ~US$35–50/admin/moTW10		US$15–25/worker/moBelow both SG incumbents. Competitive entry price + AI layer they don't have.
ACV (SMB)US$50–200K/yrUS$5–20K/yrUS$10–30K/yr10–50 workers × US$15–25/mo
Hardware COGSUS$3,000/device8N/AUS$15–30/deviceAlibaba wearable camera. 150× cheaper
Cloud/AI costIncluded in SaaSN/AUS$0.06–0.50/worker/dayGPT-4o-mini: ~US$0.00035/imgTW8; DeepSeek V2.5: ~US$0.00015/imgTW8. 400 img/day = US$1.80–4.20/worker/mo
Gross Margin75–85% (software-only)~70–80%60–75%Now fundable with corrected API pricing. Hardware cost (US$15–30) amortized over 12 months = ~US$2/mo drag.
CAC (B2B SMB)US$10–30K~US$5–10KUS$3–8KWenhao's warm network = low initial CAC
LTV (3yr)US$150–600KUS$15–60KUS$30–90K3yr × ACV. Assumes 85%+ annual retention.
LTV/CAC5–20×3–6×4–10×Healthy if retention holds
The cloud cost trap — recalculated with Feb 2026 pricing (Donna-verifiedTW8):
Previous versions of this analysis used GPT-4V pricing (US$0.01–0.03/image). That was mid-2024. As of Feb 2026, verified rates per image:

ModelCost/Imagevs GPT-4V (mid-2024)Source
GPT-4o-mini VisionUS$0.0003530–85× cheaperDonna R2 verifiedTW8
DeepSeek Vision V2.5US$0.0001567–200× cheaperDonna R2 verifiedTW8 — current price leader
Kimi K2 VisionUS$0.0002050–150× cheaperDonna R2 verifiedTW8
Claude 3.5 Haiku VisionUS$0.0004025–75× cheaperDonna R2 verifiedTW8
Trajectory: Downward curve continues. DeepSeek has cut cost >50% relative to GPT baseline. At DeepSeek rates, the unit economics improve by an additional 57%.

Updated death metric (using GPT-4o-mini baseline, conservative):
  • Always-on video: 28,800 frames/day × US$0.00035 = US$10.08/worker/day = ~US$300/worker/month. Still too expensive at US$15–25/mo revenue. Selective frames (1 per 10 sec) = ~US$30/worker/month — borderline at higher price tiers.
  • Photo per task step: 400 images/day × US$0.00035 = US$0.14/worker/day = US$4.20/worker/month. This works. At DeepSeek rates: US$0.06/day = US$1.80/worker/month — trivial.
  • Photo per room (batch QC): 20 images/day × US$0.00035 = US$0.007/worker/day = US$0.21/worker/month. Trivially cheap. Near-100% gross margin on inference.


Revised conclusion: With current API pricing, task-step photos are viable, not just batch QC. This means "Support Buddy" (Product 2) can give feedback at each step, not just post-task. The product goes from "post-task QC checker" back to "real-time buddy" — a fundamentally better product. This is the single biggest change since the original analysis. And the trajectory is our friend: if DeepSeek holds its pricing, the cost floor drops another 50%.TW8
Tailwind: inference costs are cratering. As of Feb 2026, API prices are at record lows: DeepSeek-Chat at US$0.28/M input tokens, Kimi K2 at US$0.15/M, GPT-4o-mini at US$0.15/M.TW8 For vision tasks, this trajectory means photo-based QC costs will likely drop 2–5× within 12 months. The "death metric" (inference cost > revenue per worker) becomes less dangerous with each quarter. This is the single strongest macro signal for Wenhao's unit economics.

Donna R2 signal (packing/fulfilment): @bushidoinstinct notes that digital QC is now viewed as a tool to "delete payroll and rejection simultaneously"TW5 — i.e., packing QC isn't just about catching errors, it's about removing the QC person entirely. If the AI catches errors pre-shipment, you don't need a human QC checker AND you reduce rejection/return costs. This doubles the ROI story for Wenhao's pitch.

VII. The “ChatGPT Vision” Threat

The un-asked question: "Why not just use ChatGPT?" A hotel manager can today take a photo of a room → paste into ChatGPT/Claude → get QC feedback. For free. Why pay US$15–25/worker/month?

What a dedicated product adds:

CapabilityChatGPT/ClaudeWenhao's Tool
ScaleManual — one photo at a time, copy-paste per image20 workers × 20 rooms × every day. Dashboard view across all.
WorkflowNo workflow — just a chat windowWorker snaps → AI auto-flags → manager dashboard → re-do notification
Historical dataNo memory across sessions"Room 403 flagged 3× this week" → training insight
Domain tuningGeneric AI — no hotel QC standardsHotel-specific checklists, calibrated to brand standards
Worker trackingNonePerformance trends per worker → training allocation
The moat is in the workflow, not the AI. The AI is commodity. The integration into daily operations, worker management, and QC reporting is the product. system of record, not a chatbot. If Wenhao can't articulate this clearly to buyers, ChatGPT is the real competitor.
The "SAASPOCALYPSE" signal: There's growing discourse that Anthropic's latest tools are triggering a shift from SaaS subscriptions to AI agents that do the work.TW2 Goldman Sachs is already deploying Claude for accounting/compliance — "digital co-workers" that read records, apply rules, route for approval.TW3 Implication for Wenhao: the market is moving toward AI agents that act, not just display data. His QC checker should act (auto-flag, auto-notify, auto-schedule re-inspection) — not just report. This is the difference between a dashboard and an agent.

VIII. The YC Signal — Why This Matters Now

YC Spring 2026 Request for Startups includes "AI Guidance for Physical Work" by David Lieb:5

The convergence YC identified:
1. Multimodal models can now "see" and "reason" about physics and space reliably.5
2. Hardware everywhere — phones, AirPods, smart glasses are already on workers.5
3. Labor scarcity — global shortage of skilled labor (plumbers, HVAC, nurses).5

YC's framing: "The 'Neo learning Kung Fu' moment — AI doesn't act in the world yet, but it can see, reason, and talk a human through complex physical labor."5

This is exact alignment with Wenhao's Product 2. YC RFS match means:

Live deployment signal: SAIC-GM has deployed KEPLER K2 robots at their logistics department for autonomous bin transfers — code-scanning and navigation are in-production, not demos.TW1 This validates YC's thesis: physical AI work is moving from research to deployment right now. Wenhao's human+AI approach (camera on worker, AI in cloud) is the pragmatic middle ground between "fully autonomous robot" and "no AI at all." Cheaper to deploy, easier to sell, lower liability.

IX. Playbook Dissection — Who Won and Why

CompanyModelRevenue/ScalePlaybookWhy It Doesn't Fully Apply
Augmentir6 AI connected worker SaaS ~US$10–20M ARR est. Software-first, no hardware. AI personalizes digital work instructions. Enterprise sales (US$50K+ ACV). US enterprise focus. Wenhao's market is SEA SMBs with 10–50 workers, not US factories with 500+. But the AI approach transfers.
Parsable → GE10 Connected worker, industrial ops Acquired 2023 (~US$80M raised) Mobile-first work instructions. Proved the "connected worker" category. Got absorbed into GE Digital. Validates the exit — GE paid for the category, not the scale. Wenhao's endgame could be a similar acqui-hire by Siemens, Honeywell, or Hitachi.
ServiceMax (→ PTC)14 Field service management. Scheduling, dispatch, work orders. ~US$250M+ ARR. Acquired by PTC. Rode the Salesforce ecosystem. Dominated enterprise FSM. US$10K+ ACV. Traditional SaaS — no AI, no camera, no real-time guidance. The "before" to Wenhao's "after." This is what we disrupt.
Cognex / Keyence13 Machine vision + industrial inspection. Hardware + software. Cognex: US$800M+ rev. Keyence: US$7B+ rev. Factory-floor visual inspection. Fixed cameras on production lines. Massive margins (Keyence ~55% OPM). Fixed cameras on production lines ≠ wearable cameras on mobile workers. Different use case. But validates that AI visual inspection is a proven US$8B+ market.3

X. GTM Assessment — Wenhao & Eric Contextualized

What Wenhao uniquely has

What Eric uniquely has

Recommended GTM Phases (updated with repair.sg data)

PhaseTimelineGoalWho Does What
0. Dog-food Now → Mar Deploy on Smilie's own packing team. Phone camera → cloud AI → "Did the packer pack the right items in the right arrangement?" Wenhao IS the customer — zero sales cycle. Primary question answered: does the AI actually detect errors? This is the fastest path to an answer. Wenhao: deploy + test daily on Smilie. Eric: build phone camera → AI → simple dashboard MVP.
1a. repair.sg pilot Mar → May First external pilot: Zames Chew's repair.sg team.25 He validated WTP at S$40–50/worker/mo and agreed to invest S$10K. Full-time workers = consistent usage data. Key test: does the privacy concern (camera at client premises) actually block adoption? Wenhao deploys on friend's team. Eric adapts product for field repair workflow. First external revenue.
1b. WTP validation calls Mar → May (parallel) Simultaneously run WTP conversations in volume verticals: construction safety (2 contacts, ~200 workers combined28), hotel housekeeping (2–3 contacts), commercial cleaning (2–3 companies). Don't commit to a vertical based on Smilie + repair.sg alone — both are niche. Feb 11 update: construction added as a Phase 1b target — largest market size, strongest compliance driver, but longest sales cycle. The WTP conversations will reveal whether construction main contractors actually buy from startups. Wenhao's warm network. 5–10 conversations across verticals. For construction: Wenhao's two contacts (building construction + building maintenance) are the starting point.28
2. Paying pilots May → Jul 3–5 paying pilots at US$500–2,000/mo. Include repair.sg + whichever volume vertical showed strongest pull in Phase 1b. Prove ROI: reduce QC person-hours or error rates by 30–50%. Wenhao sells. Eric iterates product. Still using phones.
3. Hardware Jul → Sep Introduce wearable camera ONLY after software PMF is proven. Alibaba sourcing for US$15–30/unit. Small batch (50–100 units). This is where hardware risk enters. Don't do this before Phase 2 validates.
4. Raise Sep → Nov YC S26 or seed round. Show 5+ paying clients, US$5–10K MRR, working prototype on wearable camera. Zames's S$10K + pilot access gives conviction signal for fundraise narrative. Wenhao pitches. YC RFS alignment5 is the narrative.
Is repair.sg really a better first pilot than Smilie? Wenhao thinks so. Here's the honest comparison:

Smilie (packing QC)repair.sg (field repair)
Sales cycleZero — Wenhao IS the customerWarm — friend, but still an external deal
WTPWenhao pays himself (no WTP test)S$40–50/worker/mo validated25
WorkersSeasonal packers, part-timeFull-time team, consistent usage
Workflow complexityLow — same task every timeHigh — varied job types
Privacy riskNone — Wenhao's own warehouseReal — camera at client premises25
Market sizeTinySmall (~5–8K technicians in SG)
FundingSelf-fundedS$10K angel

Our stance: Do both, in sequence. Smilie first (Phase 0) because it answers the hardest technical question — does the AI actually work? — with zero sales overhead. Then repair.sg (Phase 1a) because it's the first real external revenue signal and comes with angel funding. Neither is the scale vertical. The scale vertical emerges from Phase 1b WTP calls.

XI. Structural Challenges

RiskSeverityMitigation
Hardware supply chain — Wenhao flagged this as #1 risk. "Getting fked over by the factory over and over again." HIGH Delay hardware to Phase 3. Use phone cameras for Phase 1–2. When ready, start with Alibaba commodity cameras, not custom hardware.
AI processing cost — Always-on video processing is margin-negative at SMB pricing. HIGH Start with photo-based QC (snap after task completion), not video streaming. 10–50× cheaper. See death metric calculation above.
Eric's bandwidth — 10 hrs/week for hardware+software+AI is tight. MEDIUM Phase 1 MVP (phone camera → cloud AI → dashboard) is doable at 10 hrs/week for 8 weeks. But Phase 3+ needs more capacity or a third engineer.
Worker privacy / adoption — Cameras on workers raises surveillance concerns. Hotels especially sensitive. MEDIUM Frame as "Support/Safety" not "Supervision" — Donna R2 found market signals confirm adoption is significantly easier under the support framing.TW7 Process on-device or delete raw footage after AI extraction. Singapore's relatively permissive data environment helps.
Competitive timing — YC RFS5 means others will build this too. MEDIUM YC founders will target US enterprises. Wenhao's SEA SMB niche is less attractive to SV founders but more accessible to him. Geography is the moat.
Smilie exit timing — Wenhao still running Smilie. Needs to wrap up sale/transition to go all-in on this. MEDIUM Phase 0 (dog-food on Smilie packers) actually helps: the new venture starts inside the current business. Smilie becomes the first customer and test bed. Wenhao can sell Smilie while Phase 0 runs in parallel. Target: clean handoff by Q2 2026.

XII. Workflow Customization — A Manageable Problem, Not the Deepest Concern

This was initially flagged as the "deepest concern" — that every company needs custom QC workflows (repair A ≠ repair B, hotel X ≠ hotel Y, packing for client 1 ≠ client 2) and this becomes professional services, not SaaS. On closer inspection, this is a normal B2B SaaS challenge, not an existential risk. Here's why:

The actual deepest concerns, in order:
1. Does the AI actually detect errors reliably? — If GPT-4o-miniTW8 can't tell a correctly-packed box from an incorrectly-packed one from a phone photo, nothing else matters. This is testable in a weekend.
2. Will workers change behavior? — "Take a photo at each step" is a behavior change. Donna's research shows adoption is easier when framed as "Support/Safety" rather than "Supervision."TW7 But real compliance data only comes from Phase 0 on Smilie packers.
3. Will managers pay? — WTP is unknown across every vertical except packing (where Wenhao is the buyer). This is the classic sales validation question.

Workflow customization sits below all three. It's a product design challenge, not a market risk.

Workflow standardization by vertical

VerticalHow Standard?Config Needed Per CustomerScalability Risk
Packing QC Pattern is identical — "are the right items in the box in the right arrangement?" Upload reference photo + item list per SKU. Text-based config. Low — same workflow across all packing operations. Only the reference changes.
Hotel housekeeping Highly standardized — bed, bathroom, amenities, floor. Industry-wide standards. Minimal — toggle brand-specific items (e.g. "pillow at 45°"). Maybe 5–10 config fields. Low — 80% of checkpoints are identical across all hotels.
Commercial cleaning Template-based — standards defined by contract type (office, hospital, retail). Select template (office/hospital/retail) + add contract-specific items. Low — 3–5 templates cover 90% of contracts.
Logistics (pick & pack) Medium — varies by warehouse SOP, but pattern is same (correct item, correct qty). Per-warehouse item catalog + packing spec. Medium config. Medium — each warehouse has different SKUs. Item catalog management = ongoing work.
F&B kitchen prep Medium — food safety standards are regulated (SFA), but recipes differ per restaurant. Per-recipe visual standards. Potentially hundreds of items per restaurant. Medium — fast food chains (few menu items) = easy. Full-service restaurants = hard.
HVAC / repair Medium-low — varied jobs (electrical vs mechanical vs refrigerant), but each is a prompt change in 2026, not a model change. Natural-language checklist per job type. Manager writes "check refrigerant line connected, compressor seated" etc. Prompt-based config, not code. Medium — more job types than packing/hotel, but LLM-based config means onboarding is hours, not weeks. The real bottleneck is curating good checklists, which the operator (Zames) already knows.
Printing QC Medium — visual standards are clear, but per-job (color, alignment, dimensions). Upload reference + tolerance spec per print job. Medium — config is structured but high-volume. Already solved by Cognex/Keyence for large shops.
Construction safety Highly standardized — PPE requirements, scaffold inspection, housekeeping, hot-work permits are regulated by MOM WSH Act.29 80%+ of safety checks are identical across all sites. Select site type (residential, commercial, infrastructure) + toggle hazard categories. Minimal per-customer config. Low — safety checks are regulated, not discretionary. Helmet, harness, barricade, scaffolding = same everywhere. Configuration is simpler than hotel housekeeping (no brand-specific items). This is a Category 1 vertical (finish task → photo → AI checks).
The real dividing line: Verticals split into two categories:

1. "Finish task → take photo → AI checks" (simple workflow, generic):
  • Packing QC, hotel housekeeping, commercial cleaning, logistics pick/pack, construction safety
  • Same workflow everywhere. Only the QC/safety checklist changes — and that's just a text prompt, not a custom model. Construction safety is particularly standardized because the checks are regulated (WSH Act29), not discretionary.
  • This is where Wenhao should play. Config = natural language checklist. Onboarding = 30 minutes, not 30 days.
2. "Multi-step workflow with branching logic" (complex, custom):
  • HVAC repair, complex field service, multi-stage manufacturing
  • Each job type needs a different sequence of steps. Each company has different SOPs.
  • In 2019, this required custom models per workflow — hence Augmentir6 and Tulip12 raising US$45M+ and US$100M+ to build workflow engines. In 2026, each workflow variant is a text prompt to a foundation model. The gap between Category 1 and 2 is narrowing fast. An agentic system that takes an SOP document and generates a visual checklist is plausible today.
Why 2019–2023 camera startups needed custom models — and Wenhao doesn't:
Previous cycle: Agot.AI16, Dragontail19, etc. had to train custom computer vision models per use case. "Detect wrong burger" = collect 10,000 labeled burger photos → 3-month training cycle. Each new menu item = retrain.

2026 with LLM vision (GPT-4o-mini, DeepSeek Vision, Claude Haiku):TW8 Configuration is a text prompt, not a model. "Check this photo: is the bed made with corners tucked, pillows at 45°, bathroom towels folded?" That's it. No training data. No custom model. New checklist = new prompt, deployed in minutes.

This is the key architectural insight that makes "finish task → photo → AI" scalable in a way the 2019 camera startups weren't. The AI is commodity. The configuration is language. The moat is in the workflow + data layer, not the model. Google Trends confirms the shift: "Visual Inspection AI" queries are outscaling "Quality Control Software" in SEA by 3.2× as of early 2026.TW9

XIII. Research Gaps — What We Still Don't Know

This report is built on market reports, competitor analysis, founder context, and Twitter signals. The following gaps remain and should be addressed before committing resources:

GapWhy It MattersHow to ValidateOwner
Hotel housekeeping WTP We recommend this as vertical #2, but have zero willingness-to-pay data. All WTP signals are "Unknown." Wenhao talks to 2–3 hotel management contacts. Ask: "Would you pay US$15–20/housekeeper/month for AI QC?" Record specific objections. Wenhao
Commercial cleaning WTP SG has 85K cleaning workers — 20× hotels. If this vertical works, it's a much bigger business. Identify 2–3 commercial cleaning companies. Same WTP question. Are they using any digital QC today? Wenhao
Packing QC accuracy Can GPT-4o-mini reliably detect packing errors from a phone photo? We assume yes, but haven't tested. Take 20 photos of correctly vs incorrectly packed Smilie gift boxes. Run through GPT-4o-mini vision API. Measure accuracy. Eric
PSG pre-approved solutions list If competing products are already PSG-approved, that's both validation and competitive threat. Donna: search IMDA PSG pre-approved IT solutions for "field service," "quality control," "cleaning management." Donna
Worker adoption friction "Take a photo at each step" is a behavior change. Will workers actually do it? Phase 0 on Smilie packers will answer this directly. Track compliance rate. Phase 0 data
SG hotel market structure Who actually makes the QC software purchasing decision? GM? Head of Housekeeping? Procurement? Wenhao's hotel contacts can answer this. Critical for sales motion design. Wenhao
Competitor pricing Kegmil ~US$20–30/user/mo, Novade ~US$35–50/admin/mo.TW10 Our US$15–25 sits just below both while adding an AI vision layer they don't have. Donna R3 validated from Latka estimates + IMDA listing data. Donna — DONE
Vision API pricing verification ✓ CONFIRMED (Feb 7) — Donna R2 verified GPT-4o-mini at US$0.00035/img, DeepSeek V2.5 at US$0.00015/img.TW8 Feb 11 update: GPT-4o-mini pricing unchanged ($0.15/1M input tokens42). Newer models now available: gpt-4.1-nano ($0.10/1M) and gpt-5-nano ($0.05/1M)42 could be 30–65% cheaper for vision tasks. DeepSeek is now on V3.2 (text at $0.28/1M input43) — dedicated vision pricing unclear. Trajectory still strongly downward — the death metric is becoming less deadly over time. Note: OpenAI forum thread (Jul 2024) reported GPT-4o-mini uses ~33× more tokens for images vs text to equalize cost with GPT-4o — per-image costs are the same across model sizes. Always verify effective cost per image, not just per-token rate. Re-verify effective per-image cost with gpt-4.1-nano and gpt-5-nano vision. These may significantly reduce the COGS table numbers. Eric — RE-VERIFY
Camera-based QC competitor status Agot.AI locked in QSR, no hospitality expansion verified. Wobot.AI expanding to manufacturing/warehousing, not hospitality. Verkada hardware-locked in security.TW10 Donna R3 validated. Hotel housekeeping and cleaning QC remain open wedges. Donna — DONE
Construction WTP validation NEW Feb 11 Construction is the largest addressable vertical (300K+ workers27) with strongest compliance driver. But no pricing conversation yet. Wenhao's contact said "blue-collar compliance assistant" — that's interest, not a purchase order. Wenhao talks to his 2 construction contacts.28 Ask: "Would you pay US$15–25/worker/mo for AI safety monitoring? Who makes this purchasing decision — you or the main contractor?" The second question is critical — if the main contractor decides, the sales motion changes entirely.36 Wenhao
Construction site deployment feasibility NEW Feb 11 Construction sites are outdoor, dusty, wet. Can a phone/wearable camera survive and produce usable images? Worker transience (rotating crews) means repeated onboarding. Visit a site with Wenhao's contact. Take 20+ photos across conditions (height, dust, rain, bright sun). Test through vision API. Also: how often do crews rotate? Wenhao + Eric
Protex AI positioning NEW Feb 11 Protex AI (YC S21, US$36M Series B, Jan 202534) does AI safety on existing CCTV. If a construction buyer already uses or considers Protex, how does Wenhao's tool fit? Is it "instead of" or "in addition to"? Research Protex AI's pricing. Test whether their approach (fixed CCTV analysis) overlaps with wearable/phone camera guidance. Map the "facility-level vs worker-level" distinction for buyer conversations. Eric
Research priority order (updated post-Donna R3): (1) Vision API pricing — CONFIRMEDTW8. (2) Packing QC accuracy test — kills the thesis if AI can't detect errors. This is testable in a weekend. (3) WTP validation in a volume vertical — cleaning (85K workers21) or hotels (3–4.5K22) or logistics (50K21). Gift packing alone is too niche. (4) Camera competitor status — CONFIRMEDTW10 Agot locked in QSR, Wobot in manufacturing. Hotel/cleaning QC = open. (5) Competitor pricing — CONFIRMEDTW10 Kegmil ~US$20–30, Novade ~US$35–50. Our US$15–25 undercuts. (6) PSG pre-approved list — CONFIRMEDTW10 No vision QC on list = open wedge.

XIV. Red Team

The Bear Case

  • Daqri burned US$275M9 on industrial AR hardware and died. Hardware-first connected worker plays have a near-100% failure rate.
  • AI inference costs — even with corrected Feb 2026 pricing, always-on video is still margin-negative at SMB pricing. Task-step photos now work, but require worker behavior change ("take a photo at each step").
  • "Camera on worker" triggers visceral resistance from labor. Unions, workers, and even managers may reject it as surveillance. Donna R2 signal: resistance drops when framed as "Support" not "Supervision."TW7
  • Eric is 10 hrs/week in another city. Hardware+software startups need a full-time CTO. Part-time technical founders in hardware = death.
  • SMB churn — small businesses churn at 15–25%/yr on SaaS.24 With hardware in the field, churn also means device recovery logistics.
  • ChatGPT/Claude can do one-off photo QC for free. Why pay for a dedicated tool?
  • Protex AI has $36M and Fortune 500 clients.34 YC-backed, 64% risk reduction for DHL, expanding fast. They use existing CCTV — no new hardware needed. If they enter construction/SEA, they have more money, more data, and proven enterprise sales. NEW Feb 11.
  • Newmetrix couldn't scale independently — got acquired by Oracle (2022).40 Pre-acquisition had only "a handful of sales reps." If the closest analog to construction safety AI needed an Oracle acquisition to survive, what does that say about the standalone business?
  • Construction = enterprise sales. Main contractor is the buyer, not the subcontractor.36 SafetyCulture35 won construction with bottom-up PLG (free checklist app). Wenhao's product requires camera setup — that's friction that kills PLG.

Counter-Arguments (Steel-Man)

  • We're NOT doing Daqri. US$20 commodity cameras, not US$5K custom helmets. Software-first with phone cameras in Phase 1. Hardware is Phase 3.
  • API pricing has crashed 30–85× since mid-2024. GPT-4o-mini vision at ~US$0.00035/image (Donna-verifiedTW8) means task-step photos (400/day) cost only ~US$4.20/worker/month — gross margin positive. DeepSeek Vision at US$0.00015/image drops this to US$1.80/month.TW8
  • Workflow > chatbot. ChatGPT is a prompt. Our tool is a system of record — snap → AI → dashboard → historical data → worker performance.
  • Singapore labor dynamics — foreign workers in hotel/maintenance have less pushback on cameras. Employers have more control. Market signal: "Humans-as-APIs" framing emergingTW7 — workers execute tasks for AI clients. Adoption easier with "support/safety" framing, not "surveillance."
  • Wenhao is the full-time operator. Eric builds the MVP; Wenhao runs the business. Post-raise, hire a full-time engineer. Eric stays as technical advisor.
  • repair.sg = angel money + first customer in one deal.25 S$10K angel + validated WTP at S$40–50/worker/mo + full-time workers to deploy on. This is unusually strong for a pre-product venture.
  • YC alignment5 is a fundraising accelerant. "AI for Physical Work" on YC's RFS + working prototype + paying clients = strong S26 application. YC directory shows 20+ construction AI companies in recent batches44 — construction is in YC's appetite.
  • Protex AI is facility-level, not worker-level.34 They tell managers "this area is unsafe." Wenhao tells workers "your harness isn't attached." Different product, different buyer champion. They complement each other — Protex monitors the site, Wenhao guides the worker. Protex's Fortune 500 focus (DHL, UPS) is a different market from SG mid-market construction.
  • Newmetrix's failure was distribution, not product.40 Their AI worked (1.2M safety tags for Obayashi). They couldn't sell it without Oracle's enterprise distribution. Wenhao's advantage: warm contacts, SG grant ecosystem (BETC 70%31), and US$15–25/mo pricing that doesn't need enterprise sales teams.
  • Construction safety tech market is US$1.95B (2024) growing at 17.6% CAGR.45 Construction site monitoring systems: US$2.44B → US$5.13B by 2030 (16% CAGR).46 This is a bigger, faster-growing market than connected worker platforms.
  • SafetyCulture proves the GTM.35 They reached $2.7B valuation with a phone-first, free-tier, bottom-up safety inspection app. No hardware. Wenhao's Phase 1 is also phone-first. The hardware (wearable camera) comes later, after PMF is proven.

XV. Bottom Line

Verdict: Conditionally strong. Do it — but sequence correctly.

This is the opposite of Alice's AI learning app. That was a consumer play fighting ChatGPT in a saturated market. This is a B2B vertical SaaS with hardware in an underserved niche with YC validation5, warm sales leads, and a founder going all-in.

Key update (Feb 7, R2): Three critical findings shape the thesis. (1) AI inference costs have dropped 30–85× since mid-2024 — Donna-verified: GPT-4o-mini at US$0.00035/img, DeepSeek at US$0.00015/imgTW8 — making task-step photos viable. (2) Wenhao's Smilie packing problem is the strongest PMF signal (he IS the customer) but gift packing is a tiny market. The MVP vertical and the scale vertical will likely be different. (3) Workflow customization is a normal B2B SaaS challenge, not an existential threat — the real risks are AI accuracy, worker adoption, and WTP validation.

Key update (Feb 11, R3): Construction emerges as a potential scale vertical with the largest SG workforce (~300K+ workers27), strongest compliance driver (WSH Act with S$500K fines29, mandatory site cameras30), and richest grant stack (BETC 70% for SMEs31). Wenhao has 2 warm contacts (200 workers combined).28 However, construction GTM is harder than other verticals — main contractor is the buyer (longer sales cycle)36, worker crews rotate, and outdoor conditions stress hardware. No decision to commit to construction yet — it needs WTP validation like every other vertical. New competitor Protex AI (YC S21, US$36M34) addresses overlapping pain point (AI safety from cameras) but uses fixed CCTV, not wearable/worker-level guidance. SafetyCulture (US$2.7B, PLG model35) proves the safety inspection market is massive and phone-first entry works. Augmentir confirmed as category leader (Frost & Sullivan, Jan 20266b) but remains manufacturing-focused. Vision API costs continue falling with newer models (gpt-4.1-nano, gpt-5-nano42) further improving unit economics.

The minimum viable test: Phone camera → cloud AI → packing QC for Smilie's own gift packers. US$0 hardware. US$0 customer acquisition. 6 weeks build. If the packing error rate drops measurably, immediately validate WTP in larger verticals (construction: 300K+ workers27, cleaning: 85K workers21, logistics: 50K21, hotels: 3–4.5K22).

What needs to be true:

1. Smilie packing QC MVP works — AI reliably detects packing errors from phone photos (test this weekend)
2. WTP validated in at least ONE volume vertical (construction, cleaning, logistics, or hotels) — not just packing
3. Wenhao wraps Smilie sale/transition by Q2 2026 and goes full-time
4. Eric validates at 10 hrs/week through Phase 1, then commits or they hire
5. PSG pre-approval by Q3 2026 to unlock SG government subsidy (50% of costs)23
6. Apply to YC S26 with working prototype + 3–5 paying pilots in a venture-scale vertical5

New signal since last version: The repair.sg founder's dinner conversation25 is the only external WTP validation in this report. S$40–50/worker/mo is higher than our baseline estimate. Combined with S$10K angel investment (a conviction signal, not runway), this gives the venture a concrete first external pilot and an investor-customer narrative. Wenhao's instinct that repair.sg is a better first pilot than Smilie is partially right: it's a better business signal. But Smilie is a better technical test (zero friction, no privacy concerns, simpler workflow). Do both.

Why these two specifically: Wenhao knows physical ops from the inside (printing, gifting, factory relationships). He's already doing competitor research (shared kegmil.com7). Eric has built field worker onboarding systems across 4 markets at Pickupp and now builds AI agents. Eric's concern about "B2B that doesn't earn"26 is honest and solvable: Phase 0 on Smilie costs nothing, and repair.sg provides the first external revenue test. The combination of "ops-native seller" + "ops-to-tech builder" is unusually strong for a connected worker startup.

References

Market reports

[1] Grand View Research, Field Service Management Market 2024–2030 — US$6.2B (2024), 11% CAGR.
[2] MarketsandMarkets, Connected Worker Market 2024–2030 — US$5.5B → US$18B, 18% CAGR.
[3] Fortune Business Insights, AI Visual Inspection Market 2024–2030 — US$1.8B → US$8B, 28% CAGR.
[4] SEA Facilities Management Tech — estimated from regional reports + Singapore Enterprise Development Board data. Not tracked as a standalone segment by major research firms.

YC & industry signals

[5] YC Request for Startups, Spring 2026 — "AI Guidance for Physical Work" by David Lieb

Enterprise competitors

[6] Augmentir — AI-powered connected worker. US$45M+ raised. Manufacturing focus.
[7] Kegmil — SG field service SaaS. Checklists, work orders, CMMS. No AI.
[8] RealWear — Industrial smart glasses. US$180M+ raised. US$3K+/device.
[9] Daqri (defunct) — Industrial AR smart helmet. US$275M raised. Shut down 2019. Hardware-first failure.
[10] Parsable (→ GE Digital) — Connected worker for industrial ops. US$80M+ raised. Acquired 2023.
[11] SightCall — Visual assistance for field service. US$55M+ raised. AR overlays via phone/tablet.
[12] Tulip — Composable MES. Frontline operations platform. US$100M+ raised.
[13] Cognex / Keyence — Machine vision + industrial inspection. Cognex ~US$800M rev. Keyence ~US$7B rev.
[14] ServiceMax (→ PTC) — Field service management. ~US$250M+ ARR. Acquired by PTC.
[15] Novade — SG construction + facilities management. Smart forms, safety, quality.

Camera-based QC competitors

[16] Agot.AI — Kitchen QC camera for QSR. ~US$11M Series A (2023). Chick-fil-A, others. Overhead camera, custom CV models.
[17] Wobot.AI — Video intelligence for restaurant + retail + manufacturing. ~US$10M raised. AI layer on existing CCTV. India/US.
[18] Verkada — Connected security cameras + analytics. US$3.2B valuation. Security-first, ops QC is secondary.
[19] Dragontail Systems (acquired) — Kitchen management AI camera. Acquired by Yum! Brands subsidiary. Validated tech, couldn't build standalone business.
[20] Presto (PRST) — Restaurant automation, drive-thru AI. SPAC'd, struggled post-IPO. Revenue missed targets. Camera AI in noisy environments = hard.

Founder conversations & meeting notes

[25] Meeting notes: Song Fa dinner, 1 Feb 2026 — Wenhao pitched blue-collar AI idea. Zames Chew (repair.sg founder) validated WTP at S$40–50/worker/mo, would deploy to full-time workers, raised privacy concern (clients' premises). S$10K angel investment agreed (per Eric, not yet in formal notes).
[26] Eric's CRM notes, Feb 4–6 2026 — Eric: "Not advisor role — real validation." "Wants to validate + protect downside." "How to land early revenue? Eric hesitates on B2B that doesn't earn." Wenhao described as "dead set on doing this, already engineering sale of Smilie."

SG-specific data

[21] Singapore MOM Labour Market Statistics — Workforce estimates by sector: cleaning/waste mgmt (~85K), food services (~150K), transport & storage (~50K), accommodation (~15K). 2024 data.
[22] Singapore Tourism Board — Hotel Statistics — 281 gazetted hotels, ~67,000 rooms (2024).
[23] IMDA Productivity Solutions Grant (PSG) — Up to 50% support for qualifying digital solutions, capped at S$30K. EDG covers up to 50% for business upgrades.
[24] SaaS industry benchmarks — SMB annual logo churn typically 15–25% (Bessemer, OpenView SaaS benchmarks). Enterprise churn 5–10%.

Twitter/X signals — Run RT-WENHAO-202602071409 (Round 1)

[TW1] @XRoboHub — SAIC-GM KEPLER K2 deployment — Verified robot deployment at SAIC-GM logistics. Autonomous navigation + code-scanning in production.
[TW2] @olafornot — "SAASPOCALYPSE" — Signal that Anthropic tools are triggering shift from SaaS subscriptions to transactional AI agents.
[TW3] @grok — Goldman Sachs Claude deployment — GS deploying Claude AI for accounting/compliance "digital co-workers."
[TW4] @grok — API cost trajectory — Record low API prices: DeepSeek-Chat $0.28/M, Kimi K2 $0.15/M input tokens.

Twitter/X signals — Run RT-WENHAO-R2-202602071500 (Donna Round 2)

[TW5] @bushidoinstinct — Packing/fulfilment signal: digital QC viewed as tool to "delete payroll and rejection simultaneously." High conviction on packing QC ROI.
[TW6] @rentokofficial — Student housing networks moving from manual records to "real-time operations." Validates hotel/facility digital ops shift.
[TW7] @MetaBaz7 — "Humans-as-APIs" framing — workers executing tasks for AI clients. Key finding: adoption is easier when framed as "Support/Safety" not "Supervision." Critical for Wenhao's positioning.
[TW8] Donna R2 — Vision API Pricing Verification (Feb 2026) — GPT-4o-mini: $0.00035/img, DeepSeek V2.5: $0.00015/img (price leader), Kimi K2: $0.00020/img, Claude 3.5 Haiku: $0.00040/img. Downward trajectory confirmed.
[TW9] Donna R2 — Google Trends (SEA, Feb 2026) — "Visual Inspection AI" queries outscaling "Quality Control Software" by 3.2× in SEA region. AI-specific search demand is accelerating past generic QC software.
[TW10] Donna R3 — Research Gap Audit (RT-WENHAO-R3-202602071545) — PSG pre-approved list: CleanSmart, WhizWork, Infogrid (cleaning mgmt), Novade, Kegmil (field service) — no vision QC layer. Competitor pricing: Novade ~US$35–50/admin/mo, Kegmil ~US$20–30/user/mo (Latka est.). Camera QC vertical locks: Agot.AI in QSR only (no hospitality), Wobot.AI expanding to manufacturing (not hospitality), Verkada hardware-locked in security.

Construction vertical & sweep update — 11 Feb 2026 (R3)

[27] MOM Foreign Workforce Numbers (Dec 2024) — CMP (Construction, Marine, Process) work permit holders: 456,800. Construction estimated at 60–70% of CMP total (~275K–320K foreign workers). Adding local workers, total construction workforce ~300K+.
[28] Meeting notes: Wenhao & Eric, 9 Feb 2026 — Wenhao: "I got a feeling we might do the construction." Has 2 contacts: one in building construction (~100 workers), one in building maintenance (~100 workers). Contact described product as "blue-collar compliance assistant on the field." Construction interest driven by safety visibility and site shutdown prevention.
[29] MOM WSH Act — Liabilities and Penalties — Stop-work order non-compliance: up to S$500,000 fine + S$20,000/day + up to 12 months imprisonment. WSH officer mandatory for sites ≥S$10M; WSH coordinator for sites <S$10M.
[30] WSH (General Provisions) Amendment No. 2, 2024 — Effective 1 Jun 2024. Mandates video surveillance systems for construction worksites with contract value ≥S$5M. Regulatory tailwind for camera-based monitoring.
[31] BCA BETC Grant — Up to 70% for SMEs (Apr 2025–Mar 2027), 50% for non-SMEs. Covers advanced tech + manpower capability in built environment sector.
[32] MOM/WSH Council — WSH Technology Grants Resource Guide (Apr 2025) — Covers sensors, wearables, AI, AR/VR, data analytics for workplace safety. Includes CTC Grant (up to 70%), PSG, EDG, and ADS funding channels.
[33] The New Paper — SG workplace deaths (2024) — 43 workplace deaths in 2024 (up from 36 in 2023). Construction: 20 deaths (47% of total). H1 2025: 76 construction deaths + major injuries (Straits Times, Jul 2025).
[34] Protex AI — US$36M Series B (Jan 2025) — YC S21. Computer vision on existing CCTV for workplace safety. DHL: 64% risk reduction in 3 months. Clients: DHL, UPS, FedEx, GEODIS, M&S. Deloitte Fast 50 Rising Star (Dec 2025). Logistics/ports/manufacturing focus. Not in SEA or construction.
[35] SafetyCulture — AU. US$297M raised. ~US$2.7B valuation (Aug 2023). 75K+ businesses, 1.5M+ workers, 180 countries. ~US$132M revenue (2022). Bottom-up PLG: free tier → organic adoption. Digital checklists, not AI vision. Best GTM analog for safety inspection market.
[36] Verdantix — Contractor Safety Management (2025) — Main contractors are the primary buyers for construction safety tech. Centralized safety management systems deployed top-down. Subcontractors operate within main contractor's framework.
[37] SBF — SG Business Technology Adoption Survey (2024) — 73% cite high costs as #1 barrier (up from 64% in 2023). 44–47% cite upskilling challenges. PSG (67%), EDG (64%) most beneficial initiatives. 69% want more tailored digital advisory support.
[38] TechCrunch — Buildots US$45M Series D (May 2025) — Total: US$166M. US$300M valuation. AI progress tracking from 360-cameras. Clients: Turner, VINCI, Intel. Triple-digit revenue growth. 200+ employees. Tel Aviv HQ.
[39] Versatile (CraneView) — US$80M Series B (2021) — Crane-mounted AI camera. 12M+ crane picks analyzed. Monitors utilization, load, progress, safety. Tel Aviv HQ.
[40] Oracle — Newmetrix acquisition (Oct 2022) — Founded 2015 as Smartvid.io. AI safety detection from photos/video. "Vinnie" AI detects 100+ hazards. 1.2M safety tags for Obayashi. Pre-acquisition: "handful of sales reps." Now embedded in Oracle Construction Intelligence Cloud.
[41] Procore Q3 2025 Results — Revenue: US$339M (Q3 2025, +15% YoY). 80% gross margin. 17,623 customers. 2,602 customers >US$100K ARR. Public (PCOR). Dominant construction management platform.

API pricing & market data updates — 11 Feb 2026

[42] OpenAI API Pricing (Feb 2026) — GPT-4o-mini: $0.15/1M input. gpt-4.1-nano: $0.10/1M. gpt-5-nano: $0.05/1M. Note: vision tasks use ~33× more tokens than text (structural, not a bug — confirmed Jul 2024 forum thread + OpenAI DX head). Effective per-image cost unchanged vs text per-token rate.
[43] DeepSeek API Pricing (Feb 2026) — DeepSeek V3.2: $0.28/1M input (cache miss), $0.028/1M (cache hit). No dedicated vision pricing published. Report's earlier $0.00015/img figure was for V2.5 (Feb 2025 pricing).

Construction safety market — 11 Feb 2026

[44] YC Company Directory search (Feb 2026) — 20+ construction AI companies in recent batches: Fresco (F24), Karmen (F24), Structured AI (F25), Bild AI (W25), SmartSite (S16), Opusense AI (Sp25). Construction is actively in YC's investment appetite.
[45] Dataintelo — Construction Safety AI Market — US$1.95B (2024), 17.6% CAGR, projected US$9.17B by 2033.
[46] GlobeNewsWire — Construction Site Monitoring Systems Market (Jan 2026) — US$2.44B (2025) → US$5.13B (2030), 16% CAGR. Driven by AI analytics, cloud platforms, drones.

Competitor updates — 11 Feb 2026

[6b] Augmentir — Expansion & Product Updates (Mar 2025) — 38% team growth H2 2024. New clients: Colgate-Palmolive, Mondelēz, Duracell, Hitachi Energy. Frost & Sullivan leader (Jan 2026). 5M+ AI time/motion studies. AR features (Sep 2025). Video-to-procedure GenAI (Mar 2025).
[7b] Kegmil — SGD$2.2M Pre-Series A — Led by ME Innovation Fund + Origgin Capital. Expanding into construction, manufacturing, marine offshore, renewables. PSG pre-approved. SG-based.