I take products from an idea to something people actually use.
A boutique problem-solving studio: software, AI automation, product and growth, with a small network of specialists. I run it and I still write most of the code.
Lead engineer on a live-streaming platform for a media client, built end to end and deployed on Vercel: stream coordination, the products used during live streams, and the content production around them.
Technology decisions on new and existing projects and SaaS products, task management, coordinating the tech team and being the link between management and engineering — while still coding, deploying and doing QA.
Managed the whole company: people, projects and clients. Learned a lot of tools and processes, and how to run a team that ships.
Most of my work now is LLM-backed systems: agent orchestration with LangGraph, retrieval pipelines and vector search, real-time voice, and the token accounting, provider abstraction and healing loops that keep all of it affordable and alive in production.
The whole product around the model: Next.js and React on the front, Node, Express and Hono behind it, Expo when it needs to be in a pocket, tests and CI so it stays shipped.
Started at twelve and never stopped. Tailwind, SCSS, motion, accessibility, and the performance work that makes a site feel instant.
Unreal Engine 5 with blueprints, Blender, and Python for the pipelines around them — a shipped game and a 10,000-character generative collection.
Six years as a full-time creator: trailers, voice-overs, edits, streams, and the marketing that grew two audiences from zero.
The unglamorous layer that keeps products up: Linux, Docker, CI, deployments on Vercel and Railway, headless CMSes, WordPress, and a pentester's toolkit.
LLM-backed systems that hold up outside a demo: provider abstraction, streaming, function calling, telemetry, and sane behaviour when a model or an API decides to fail.
Multi-step agents that actually finish the job: LangGraph orchestration, tool use, queue-backed workers, human-in-the-loop approvals, and healing loops fed by real error telemetry.
Getting the right context in front of the model: ingestion and chunking pipelines, vector search, hybrid retrieval and ranking, and the plumbing that keeps the index fresh.
Real-time voice that survives a real room: voice activity detection, echo cancellation, barge-in, wake words, diarization and continuous conversation instead of push-to-talk.
Knowing what every call costs before the invoice tells you: per-provider token counting, cost matrices, metering middleware, wallets and ledgers, and usage-based billing on Stripe or PayPal.
The product around the model: Next.js front-ends, Node/Hono services, Postgres/Supabase and Mongo, mobile with Expo, tests, CI and deployment on Vercel or Railway.
I've been building things on the internet since 2010. Websites, apps, platforms, streams, audiences — usually several at once, because I've never been able to leave a project at "it works" when it could also be found, used, and paid for.
That's the short version of what I do: I take a product from an idea to something people actually use.
