Projects I've actually
shipped.

Multi-agent system that turns plain-English instructions into self-healing Playwright tests, executed at scale on cloud browsers.

Planning, generation, execution, validation, and self-healing run as a LangGraph multi-agent loop over Browserbase and Stagehand. Tests survive selector drift without manual edits.

AI compliance layer between regulated companies and their regulators — validates filings, benchmarks disclosures against peers, and runs end-to-end filing prep with humans in the loop.

Multi-agent orchestration with LangGraph and Pydantic AI, grounded in millions of SEC filings. Ships as a Word add-in and Outlook plugin, SOC 2 Type II.

Revolutionised grading for engineering CAD work — saving professors thousands of hours at Imperial College London, UCL, and 40+ institutions worldwide.

First engineering hire at a VC-backed startup. Led the ground-up rebuild on AWS serverless with TypeScript and Next.js. p95 3.1s → 680ms; grading engine scales to thousands of concurrent submissions.

Mental-health platform pairing licensed therapists with anonymous audio rooms — confidential by default, personalised by design.

Audio rooms via Stream/WebRTC, session matching, and mood tracking on a Supabase + OpenAI stack. Launched with licensed therapists across multiple time zones.

NFT marketplace for creators to mint, buy, and sell digital collectibles, with wallet auth and on-chain transactions.

React + thirdweb SDK on the frontend, Sanity CMS for metadata and collections. Supported multiple wallets and on-chain minting flows.

Bank for the underbanked. iOS, Android, and an admin console, shipped as a lean team of four.

One React Native codebase served iOS and Android; the admin tool ran the same design system on web. Reached 73k customers inside a year.

Online learning for Nigerian universities. Grew from MVP to funded startup with an active user base.

Built the full product from scratch — live sessions, assignments, payments, admin. Accepted into top incubators; grew MAU 3× in six months.

Real-time face mask detection for safety compliance, tuned for low-power edge deployment.

Custom CNN on MobileNetV2, optimised with TensorFlow Lite for Raspberry Pi. 30+ FPS on modest hardware with 95%+ accuracy.

Natural-language query layer over Revit BIM models, letting engineers pull structural and material data without touching the UI.

Python + Revit API + ODBC, with an NLP front door so engineers could ask questions and get reports without navigating complex Revit screens.