AI products, growth systems and technical prototypes.

Outcome

2,280+ organic installs with no paid acquisition. The MCP server, Vault (paid cloud sync with paying customers) and a live PWA are all shipped.

Key decisions

• Treated privacy as the product: all data stays on device, which shaped the entire local retrieval architecture rather than being a settings toggle.
• Shipped an MCP server so the same local index is usable from coding agents and LLM tools, not only the extension UI.
• Monetised with Vault, a paid end-to-end encrypted cloud sync tier, rather than ads or data monetisation.
• Built LLMnesia Insights, a supporting agentic analytics pipeline: PostHog, GA4 and Search Console evidence is collected into Supabase, then a coding agent reasons over the evidence to publish structured weekly product reports.
• Closed the loop on distribution with an SEO and LLM-discovery site (MDX content, sitemap, feed, llms.txt) to drive organic installs.

Outcome

Reached 12,000+ users with zero paid acquisition and sustained it without an acquisition budget.

Key decisions

• Built around the writing workflows teachers and admins repeat, rather than a general-purpose chat interface.
• Scaled through SEO, organic content and directory listings instead of advertising spend, so every improvement had to earn real usage.

Outcome

Working MVP covering the full loop: intake questionnaire, AI-generated plan, daily sleep diary, weekly AI reviews, living plan updates, PDF export and paid access.

Key decisions

• Grounded generated plans in an AI-structured knowledge base of 73,000 words, organised by baby age, sleep problem and training method, with a loader that selects only the relevant knowledge for each intake.
• Designed the plan as a living document: parents log a daily sleep diary, a weekly AI review runs against it, and the plan updates instead of going stale.
• Designed the prompt and output structure around practical, parent-readable plans, not raw model output.
• Kept the stack conventional: Next.js, Supabase and Stripe over the Gemini API, so build time went into product rather than plumbing.

Outcome

A working end-to-end demo: a non-upgradeable USDC treasury contract deployed on Base Sepolia, a credits gateway and dashboard on Vercel, and a vendor demo service on DigitalOcean, all wired into a scripted 402, top-up and retry flow.

Key decisions

• Designed the flow so agents pay in credits while USDC settles on-chain: a 402 response triggers a top-up and a retry, so payment never breaks the API call.
• Wrote and deployed the on-chain treasury escrow: USDC deposits via EIP-2612 permit and Permit2, per-user spending limits enforced in the contract, gateway-only batch settlement, re-entrancy guards and Foundry test coverage.
• Built the off-chain side end to end: a Next.js gateway and credits dashboard with HMAC-signed vendor requests, and an Express vendor demo that renders the full request chain with inline trace logs.
• Tested the product story first with a static marketing site and private beta interest capture before building deeper.

Outcome

Prototype stage, not launched. Early exploration of how much hidden outlet context can be surfaced at read time.

Key decisions

• Scoped it as a prototype: validate that outlet ownership and donation data can be surfaced reliably before investing in a launch.
• Surfaces context where the article already is, rather than sending readers to a separate lookup.

Outcome

Ongoing founder coaching covering strategy, prioritisation and execution.

Key decisions

• Regular sessions centred on priorities, decisions and follow-through rather than deliverables.
• Clarified where AI should sit in the product rather than bolting it on as a feature.