I’ve spent ten years turning the unclear kind into the boring kind — and boring ships. Teams call me in when something feels wrong but nobody can name it yet: the roadmap that keeps slipping, the “quick” AI feature that never stabilises, the codebase everyone’s afraid to touch. My job is to make it clear. Then make it boring.
How I think
When a team’s language is muddled — when “user” quietly means five different things — the boundaries are muddled too, and no framework will save you. So I fix the language first. The architecture tends to fall out of it.
Then I write down the trade-offs and the failure modes before a line of code, and choose the smallest design that survives the next twelve months — not the cleverest one. I build it typed, tested, and documented well enough that you own it after I’m gone. Cleverness is a liability; I optimise for whoever reads this at 2 a.m., six months from now.
Field notes
The same expensive mistakes, over and over. A handful I’ll die on:
Your architecture problem is usually a naming problem.
Muddled words make muddled boundaries. Get the language right and the diagram writes itself; get it wrong and no amount of microservices will save you.
Adding AI is not a strategy.
Most “AI features” are a demo in a trench coat — dazzling in the meeting, unmaintainable by Friday. I start from the job to be done and the evals. The model is the easy part; the retrieval, guardrails and fallbacks are the work.
The unglamorous work is the leverage.
Test harnesses, CI, internal tooling — they never demo well. They just quietly make a team twice as fast and far less afraid to ship. I build the boring infrastructure that pays for itself in a month.
You don’t have a scale problem yet.
Most systems get re-architected for a load they’ll never see, at the cost of the shipping they need today. Design for one more order of magnitude, not six. Buy the scale you need with the money you have.
Adding people to a late project makes it later.
The bottleneck is rarely hands; it’s unclear ownership and slow feedback loops. I fix the loop before I’d ever reach for more headcount.
Evidence
Three things I own end to end. Most of the rest is under NDA — so instead of a gallery, here’s proof of the thinking above.
Proof I take a fuzzy idea — “capture leads with AI” — to a shipped, secure, end-to-end system.
A Chrome extension captures the lead on the page you’re viewing; an AI backend scores fit and drafts outreach with Claude; a Next.js dashboard makes every past lead searchable by meaning with pgvector RAG. Secrets stay server-side; the extension ships with minimal permissions.
Proof I ship real products people use — not demos.
A free, no-sign-up tool that turns any image into AI-ready prompts for Midjourney, Stable Diffusion and Flux — six modes, structured output, multiple languages, vision-AI powered. Fifteen generations a day, no card, no watermarks.
Proof a good pattern pays for itself: I extracted the RAG pipeline from Exhibit A and shipped a second live product from it — grounded, cited, and private by design.
Paste or upload a contract and get a plain-language summary, the clauses worth reviewing, and answers grounded in the text — each one citing the exact clause it came from. When the document doesn’t support an answer it says so instead of guessing. The document never touches the server: parsing and retrieval run per request and the index lives only in your browser tab.
Engagements
The lowest-risk way to start is a fixed-scope paid review: a few days, a clear written deliverable, and an easy decision about what to do next. Easy to scale up from there.
Let’s talk
A review, an AI integration, a workflow audit, or a full build — tell me the problem and the outcome you’re after. The easiest place to start is a fixed-scope paid review: a few days, a clear deliverable, low risk. I usually reply within a day.
Currently taking 1–2 new clients
