AI Product Engineer

I'm a founding product person — I do the product judgment and I build the thing, from "is this worth solving" to systems built with real test suites behind them.

founding-team, zero-to-one · designs, builds, and evaluates AI products · dual US/EU

Case studies

Engineering rigor

Chess trainer

A training tool that helps players study by checking their moves against real engine analysis rather than a set of hand-written rules.

The decision I chose to validate against Stockfish rather than heuristics — objective truth over plausible-looking output. Move quality is not something a rulebook should be guessing at when a real engine can just answer the question.

Proof 236 automated tests (as of 2026-07-24). The app itself validates every move against a real Stockfish binary — that's the architectural decision. The suite that checks the app is contract-tested against a fake engine implementing the same interface, with expected values (golden fixtures) seeded from genuine Stockfish runs, plus one test that spins up the real engine end-to-end.

What it shows I build things that can be checked, not just demoed.

Read the full case →

AI judgment

OIS — opportunity gate

An LLM gate I built and ran in an n8n pipeline (built, not running now) — deciding which raw candidates are worth a closer look and which get discarded early.

The decision The gate's job was coarse sieving, not quality ranking — so the metric that matters is false-drop rate, not accuracy. A gate that quietly discards good signal fails invisibly, and accuracy alone would hide exactly that failure.

Proof There is no eval number yet, and I won't substitute accuracy for one: I have not measured this gate's false-drop rate. The measurement is specified, and a reusable eval harness is the next build, precisely so this gets measured rather than estimated.

What it shows I reason about which failure mode matters before optimizing.

Read the full case →

Past project

ProntoVA

A multi-tenant WhatsApp + LLM platform for appointment-based clinics, built end to end by me. I stopped it on 2026-09-03; it is no longer running.

The decision I picked a narrow buyer rather than a general chatbot framework, then ended the project before launch pending a data-protection review — rather than let it drift.

Proof Built and exercised end to end: one backend running many bots, each defined by a config file, with swappable channel, model, payment and speech adapters, tenancy-isolation tests, per-call cost metering and a GDPR privacy layer. It never reached a customer.

What it shows I can build a whole AI product, and I stop one on purpose when it can't go forward.

Read the full case →

How I work

  1. Pick the problem Narrow the buyer before writing code — ProntoVA (a past project) targeted appointment-based clinics, not "AI for SMBs".
  2. Design it Decide which failure mode matters first — the OIS gate was judged on false-drop rate, because a gate that discards good signal fails invisibly.
  3. Build it Make correctness checkable — the chess trainer's suite is contract-tested against a fake engine with fixtures seeded from real Stockfish runs.
  4. Measure whether it worked Say plainly what is not measured yet — the OIS gate has no false-drop number, which is why the eval harness is next.

Stack

Languages & runtime

TypeScript, Node.js, Python

AI & automation

Claude & LLM APIs, n8n

Data & infra

Postgres, Supabase, Redis, Docker

Integrations & quality

WhatsApp Business API, pytest, CI ratchet discipline

Contact

Dual US/EU work authorization.