I build the services that take untrusted input — and the models that know what it means.

A React SPA on Cloudflare Pages talks to a FastAPI service on Fly backed by SQLite in WAL mode; a separate TTS worker fleet runs kokoro-onnx. The worker never touches the database — it claims and completes jobs through internal endpoints over Fly's private network, so SQLite keeps exactly one writer. Quota is transactional: the free tier's three lifetime conversions are checked and incremented inside one transaction, and refunded when a job permanently fails. The queue is crash-safe — jobs running past 15 minutes are requeued, three failed attempts mark the job failed with a user-visible reason and a retry. Synthesis sits behind a single synthesize() interface, so a premium voice provider drops in as a higher tier without touching the queue. Magic-link and Google auth, Stripe webhooks for billing.

Fastify and TypeScript over Prisma, PostgreSQL and Redis, with Zod at every boundary; the client is Flutter with Riverpod, GoRouter and flutter_tts. The AI layer is provider-abstracted across OpenAI and Anthropic and runs a hybrid rules-plus-LLM path with output validators, so a model swap doesn't change behaviour. Five MVP modules: a microtask engine that splits a goal into 3–12 atomic steps, reading support with an original/simplified toggle and line focus, timed focus sessions, an explicitly user-controlled preference profile that never infers a diagnosis, and first-party analytics the user can export and hard-delete. Docker Compose, GitHub Actions, nginx with Let's Encrypt, Prometheus.

Split across two repos. The web app is React with React Three Fiber and drei for the 3D scene, Zustand for local state, TanStack Query for server state, and Framer Motion for transitions. The API is Fastify with JWT in httpOnly cookies, bcrypt for credentials, Zod schemas on every route, and PostgreSQL over node-postgres.

Attendance is validated server-side against the event's coordinates rather than trusted from the device, so a check-in is only accepted inside the event's radius during its window. Express handles the REST surface; the Flutter app covers event discovery, check-in and tracking.

Flask API over MongoDB Atlas with authentication and per-user profiles; the editor renders a live preview as you type and exports through jsPDF client-side, so document generation never blocks a request. Responsive down to phone widths.

Conversion runs entirely client-side, so documents never leave the browser. CI is the interesting half: every push to main runs lint and build, publishes dist to the gh-pages branch only when the build passes, and pull requests run the same checks without deploying.