Deep Market Research

Identity technologies reach universality. Tool technologies plateau. Personal agents (persistent, contextual, representing you) are identity — not tools. 10–50K runners today = ARPANET-era email. The generic “AI agent” market is 90% agentwashing (130 genuine out of thousands). Ignore that noise. The personal agent adoption curve follows email/smartphone, not VR/IoT. Hotmail moment ~2027–2029. Build for the 10–50K early adopters now.

Do it — but sequence correctly. Construction emerges as largest potential scale vertical (300K+ workers, WSH compliance, BETC 70% grants). Protex AI ($36M) is a new competitor but uses fixed CCTV, not worker-level guidance. SafetyCulture ($2.7B) proves PLG safety inspection market. Sequencing unchanged: packing QC dog-food → repair.sg pilot → validate construction + cleaning + hotels in parallel.

Flag-planting side project, not primary focus. MCP standardisation + agent proliferation = clear demand for quality/routing layer. Timing is early — not enough agents to rank yet. Eric is building agents (Donna, avet, OpenClaw) = natural supply side.

Fills the widest gap in the agent infrastructure stack. Academic benchmarks serve researchers. DevOps tools serve engineers. Nobody serves end-users asking “is this agent good?” HubSpot Website Grader drove 10M+ leads. 92%+ margin, $0.05/eval COGS. Natural eval layer for Agent Elo. Start with objective dimensions only.

Start with manual concierge, not a platform. SetupClaw ($20k/wk) and 182K GitHub stars prove demand. Chat-triggered UX is genuinely novel. But ClawCloud Run (official ecosystem platform) could add 1-click OpenClaw deploy anytime. Test with 10 manual setups at $50–150 before building. Build the platform only after 50 manual setups.

Viable local SaaS — but wrong deal structure. ~50% gross margin, HKD $100–150K/school, ~HKD 2–3M ARR at 20 schools. HK$500M govt fund tailwind. LingoTask 7.5x ahead + QEF free. Hourly beats profit-share 1.6–3x. Best: blended deal (lower hourly + rev share). 4-week prototype sprint first.

Worth experimenting — but validate WhatsApp leads first. Existing ola.tech business (20K+ customers, inbound leads) + Raymond in Shenzhen (Huaqiangbei sourcing) + Eric’s WA agent architecture = structurally attractive. Near-zero CAC + near-zero ops cost + world’s best sourcing. But zero phone-trading experience and working capital intensity are real risks. 2-week manual test costs $0.

Real gap, wrong shape for standalone business today. Agent infra stack has funded players for compute (E2B $32M), browser (Browserbase $300M), email (AgentMail YC), payments (x402) — but nobody owns phone. Gap is real. But phone is centralized (carriers + Meta), Twilio can absorb in one sprint, and market is 10–50K runners. Ship open-source MCP server (10 hrs), bundle with OpenClaw hosting. Don’t build a company.

Upgraded from weekend hack to phased experiment worth pursuing on Discord. Broadened thesis — community gating is real pain. 95% gross margin on Discord/TG. Start with bot, prove retention lift, then expand. WhatsApp is a margin trap — avoid until unit economics proven elsewhere.

Conditionally interesting. High ceiling, fragile floor. 10+ gesture ring companies failed (Nod, Logbar, Ringly, Motiv). Vibe-coding angle is new but hardware supply chain risk is real for solo founder. No recurring revenue = structurally fragile unit economics.

Don’t build a standalone directory. 100+ AI directories exist. Space shows 2000s web directory crash patterns. No dog-food signal. Builder profiles emerge naturally inside Agent Elo when builders submit agents for ranking. The directory is a feature, not a product.

Thesis reframed: dropped yoga (lost strongest moat), now music head-on with 30+ schools. Value prop cleaner (“sibling hub”) but competitive position weaker. 700 sqft must split into 2 concurrent rooms. Break-even tighter (42–47% util). Validate with pop-ups: do parents actually bring BOTH kids?

Macro thesis is real, but “trade domains” doesn’t follow. Most domainers lose money. Platforms consolidated (GoDaddy $437M). Tokenized domains unproven (ENS: 22 txns/day). Agent identity infra being built by trillion-dollar companies. Only viable play: productize existing CLI as SaaS for AI builders. Side experiment at best.

Consumer app: skip. No distribution, no ad budget, no growth skills. LAIX tried with US$100M and got delisted. Realistic consumer LTV:CAC is negative. B2B corporate AI readiness using Alice’s HR contacts is the only viable path — US$72–240K ARR Year 1 with zero ad spend.