Hey! I'm
I build and ship production software — fast. A senior full-stack engineer with an AI-accelerated workflow that turns weeks of delivery into days.
I founded and led a software company — a team of ~15 engineers at peak. With a modern stack and an AI-accelerated workflow, I now deliver that same scope with 2–3 people. I've run it both ways.
~15 engineers→same scope→2–3 people
Available for contract and fixed-scope work — remote, across EU and US hours.
// what I sell
An agreed scope at an agreed price, working and deployed in days rather than months.
LLMs and automation inside a product people actually use, not a demo that impresses once.
Making the systems a business already runs on talk to each other, and stay in sync.
Senior engineering judgment on the days you need it, without a permanent hire.
// track record
~15 → 2–3engineers for the same delivery scope
18 yrsshipping production software, since 2008
100+projects delivered
1000+features shipped
2008–2018
Grew it from the first paying client to a ~15-engineer team, owning architecture, hiring and delivery across the whole portfolio — and the client relationships that funded it. Ran the business side and the codebase at the same time.
2018–now
I own the architecture and the product. Every feature in it began as a flow I sketched and a decision I made about how it should work — and I write a large part of it myself. Multi-tenant across 27 locations in 9 countries: offline-first field apps, AI document pipelines, computer-vision workflows, SSO and permissions, and the integrations that keep separate systems in sync. Production, not prototypes.
2020–now
Fixed-scope builds, AI integration sprints and fractional tech-lead work, taken on directly. Small scopes, agreed up front, shipped.
// proof
A civic map of small neglected places, live and open source.
No accounts at all — a device-generated id is the whole identity model, which is what shaped rate limiting, moderation and abuse handling. The design problem is proving a place was improved when you know nothing about who says so.
Cloudflare Workers · D1 · R2 · Vue 3 PWA · 288 tests
Nine sources swept nightly, scored by an LLM in two lanes, one ranked digest.
Scoring one item per model call ate 35 of the platform's 50-subrequest budget, so the scorer — not the source list — was capping the funnel. Batching took it past 200 a night without adding a source.
serverless cron · Claude API · KV · 133 tests
// and the shapes I build most often
capture→local queue→reconcile→sync
A durable local queue, ordered replay, and conflicts resolved against server state on reconnect — so nobody has to think about being offline.
TypeScript · IndexedDB · service workers · REST / JWT
ingest→extract→classify→validate→report
Isolated stages, each retrying on its own, with an audit record per step. One bad input never stalls the batch.
Python · serverless · Claude API · OCR
image→on-device model→vision-LLM fallback→structured result
A fast on-device model locates and segments; anything it is unsure about escalates to a vision-LLM. Pixels in, checkable data out.
YOLO · on-device CV · vision LLM · TypeScript
pull sources→narrow to active→aggregate→generate→schedule
Pulls several internal APIs, narrows to the accounts that actually moved, and says plainly when nothing did.
Python · scheduled jobs · REST APIs · retry/backoff
reference prices→match items→price→snapshot→serve
Prices a set of items against a reference table and serves it — with the snapshot built free of personal data.
Python · FastAPI · SQL · pytest
source API→transform→dedup + retry→destination
Small bridges pulling from one API and pushing to another: dedup, retries, and a mock of the far end so failures are testable.
serverless · AWS Lambda · Terraform · queues · REST / SOAP · webhooks
// also, a lot of
// how I work
talk to users→requirements→sketch workflow→align→prototype→deploy→real use→tune
I talk to the people who'll actually use the thing, agree the workflow before anyone writes code, then get a functional prototype deployed and in real use as fast as possible. Everything after that is tuning against reality instead of guessing — which is why scope stays honest and nothing gets built for six months before someone notices it's wrong. Tasks and use cases get documented as I go.
// stack
// say hey
Got something that needs building?
→[email protected]Usually replies the same day. Happy to start with a 20-minute call.