I build production AI products, and the agentic infrastructure under them.

Scheduling SaaS built end to end: a client booking UI, a native iOS app, Stripe payments, transactional email, and admin endpoints. Every piece of production plumbing owned alone.

• The schedule engine is a layered, composable pipeline: weekly recurring rules, then date overrides and blackouts, subtract existing bookings, apply min-notice and slot-granularity constraints. New rule types extend it without a rewrite.
• Solo on all of it, Stripe, auth, Resend, Cloudflare D1/R2/Workers. The iOS client is native SwiftUI.
• Wrapped in an AI-driven ops loop: living PRDs stay in sync with shipped code, autonomous QA runs like a human, automatic maintainance, and a weekly job proposes what to build next.

Built AI-native with real engineering discipline: Claude Design for interface, Claude Code for implementation, structured agent context, code review, and an e2e test gate. This is AI-agent-driven development plus product-scoping judgment.

Stack

TypeScript/TanStack/React/Cloudflare/PostgreSQL/Swift & SwiftUI/Stripe/Resend

A Model Context Protocol server that lets Claude answer natural-language questions against a private PostgreSQL database behind real auth, e.g. "who owes me" over a personal ledger.

• OAuth 2.1 end to end, dynamic client registration, PKCE, cookie sessions. I implemented it rather than lean on a managed provider, keeping full control of credentials and exactly what the model can reach.
• Scoped the model to a read-only DB role: it can read, but cannot write, drop, or reach other schemas, so a prompt-injection attempt has nothing to exfiltrate or mutate.
• Deployed on a personal Proxmox homelab behind a Cloudflare Tunnel as an OpenRC service.

Stack

TypeScript/Model Context Protocol SDK/OAuth 2.1/PostgreSQL/Cloudflare Tunnel/Proxmox & LXC

An MCP server that gives an agent a hands-free end-to-end browser: navigate, click, type, screenshot, and assert against a live app.

• Passes solid, reusable scripts to the browser-control layer (Chrome DevTools Protocol) rather than narrating each step, a deliberately token-efficient design.
• Runs as the agent-eval loop for Satisfait Là: drives real user flows and verifies behavior, closing the retry loop on AI-generated UI changes before they ship.
• Complements deterministic Playwright e2e (the PR gate) with the human, visual, and UX checks code assertions miss, returning screenshots as evidence. Built around layered, maintained context documents the agent runs against.

Stack

TypeScript/Model Context Protocol SDK/Chrome DevTools Protocol/Proxmox & LXC

Web across the product: the customer-facing website, in-browser tooling for partners, and Bluetooth-connected web apps talking to physical devices.

• Built the customer-facing product website in React on a headless CMS (DatoCMS) with a GraphQL data layer — first on Gatsby, later Next.js with Tailwind.
• Built the Node/Express integration layer: the REST and GraphQL endpoints wiring mobile apps, partner tooling (retailer device orders, oil-filler intake), and the headless CMS to the backend's Java Spring Boot microservices — worked across that Spring Boot backend at a working level, not as its owner.
• Built an in-browser batch QR scanner so partners could scan a full tray of 50 pods at once from the web app. Codes were 15 mm × 15 mm under inconsistent lighting; ZXing couldn't read them, so I integrated OpenCV's pre-trained WeChat QR model to decode where the standard library failed.
• Built Web Bluetooth apps with real-time device state and paired-device sessions, resolving vendor quirks at the protocol level so behavior matched across Chrome, iOS WebView, and Android WebView.
• Built and maintained the Airgraft Merch store on Shopify (Liquid theme/template system).

The QR work integrated a pre-trained model into the pipeline — I didn't train anything, and I don't claim a decode rate. The honest ceiling: reading the full tray where ZXing failed.

Stack

TypeScript/React/Next.js & Gatsby/Tailwind CSS/DatoCMS/GraphQL/Node & Express/Web Bluetooth API/OpenCV (WeChat QR)/Shopify Liquid

An embedded engagement inside a large US enterprise client's own team and infrastructure, working directly with their engineers to take a migration from a vague spec to a production pipeline.

• Built the front end and migration pipeline for an LLM-driven ~700 MiB data migration from WordPress (Gutenberg blocks) to Contentful: ~150 Gutenberg block types mapped into ~50 Contentful content models, defined as the migration progressed.
• The source was too irregular for a rigid deterministic script, so I routed the ambiguous mappings through an LLM path with schema validation.
• Built a version-control layer over populated data (compare, roll back, re-import without losing reviewer work), a state machine over import steps, and a human-in-the-loop QA review tool that turned reviewer notes into approvable correction actions.

Stack

TypeScript & JavaScript/OpenAI + Gemini APIs/Contentful (GraphQL)

Tooling I've open-sourced around agentic development and self-hosted infrastructure.

• end-to-end-review-skills — Agent Skills for layer-by-layer code review in Claude Code (frontend, backend, database, security) with an orchestrator that merges them into one report.
• homelab-cloud-browser — a self-hosted Chrome DevTools Protocol browser endpoint on a Proxmox LXC: headless Chromium plus an auth proxy, a drop-in browser backend for agent tooling.