AI EngineerML Engineer
Building production AI systems, and researching how agents learn from generated worlds.
A question arrives, with the tenant and language it came in.
One production AI platform, part by part. Pick one to see what it does and which case study it belongs to.
8 parts of one platform. Pick one.
Systems running for real users at Bizzbee. Diagrams first; the method is one click away.
where is order 84-5512? — An identifier has nothing to embed. Lexical search finds it exactly; the vector arm returns other order enquiries.
Built retrieval as an agent rather than a fixed pipeline: it decides whether to retrieve at all, which strategy fits the question, and whether what came back is enough to answer with.
Extracted entities and relations from the corpus into a graph, so questions that need several facts connected can be answered by traversal rather than similarity.
Nothing identifying, nothing unsafe. The response goes out unchanged.
Built moderation and personal-data anonymisation into the output path itself, with constraints configurable per deployment.
Built ingestion and chunking pipelines for structured and unstructured sources, shaping chunk boundaries so retrieved context stays coherent.
WITHOUT MEASUREMENT / WITH
CHANGE → ?
A prompt or retriever is edited. Behaviour moves. Nobody can say in which direction, or whether to keep it.
CHANGE → SCORE → KEEP / REVERT
The same change is scored against a fixed set before it ships, and production feedback goes back into the set.
Built evaluation pipelines and feedback loops that quantify model behaviour in production, so each iteration is judged rather than guessed.
No performance figures appear on these four. The work is under NDA-shaped constraints and nothing was measured into a number I can publish and stand behind. The measured results are in the open-source work below, where anyone can reproduce them.
Built outside work, deployed, and measured. Both publish what they found — including where the result went against me.
GENERATED ON-DEVICE · THE ELEVEN THAT SCORED HIGHEST WHEN RATED BLIND
WHAT EACH MODEL SAYS THE CHURN RATE IS, AGAINST 20.4% WHO ACTUALLY LEFT
The line is the rate that actually occurred, at 20.4%.
Trained a churn model, checked whether its probabilities mean what they say, then derived the decision threshold from what a retention offer costs rather than defaulting to 0.5.
SHARE OF A/A EXPERIMENTS SHOWING A FALSE RESULT, BY HOW OFTEN THEY WERE CHECKED
The line is the error rate the test promises, at 5.0%.
Audited a real mobile-game experiment with 90,189 players: the significance test, the effect sizes it was equipped to find, and what watching the dashboard would have done to the answer.
MEAN HUMAN RATING OF EACH FILTER'S TOP TEN, OUT OF 5
The line is no filter at all, at 2.23.
Generated sixty ad creatives, scored them four ways automatically, rated all sixty blind, and measured whether the scores agree with the human.
ANNUAL REPORTS PARSED WITH NO WRONG CONTENT, OUT OF 20
The line is the regex parser it replaced, at 14.
A LangGraph agent that turns a ticker into a research note where every claim carries the filing passage behind it and every figure comes from the company's own filed data.
M.Sc. thesis direction, and the chain of work leading into it. Early research — there are no results here yet.
A learned simulator to plan inside.
If the simulator is good enough, planning no longer needs the real environment.
THE OPEN QUESTION
A world model that generates plausible trajectories lets a policy be trained and tested without acting in the real environment — which matters most exactly where acting to learn is expensive or unsafe. The part I care about is constraint: a generated trajectory is only useful if it could actually have happened.
ts-worldmodel-lab — benchmarking harnessBizzbee
Production LLM systems for a multi-tenant AI customer service platform: retrieval, generation, safety and the evaluation loops around them.
CORE
AI / ML
DATA
INFRASTRUCTURE
Open to AI engineering, ML engineering and data science roles.