MSc Artificial Intelligence, NCI · graduating January 2027
One email. I reply with two times that work.
Every system below names the failure it defends against before the thing it adds. That order is the argument, and it is the reason this site has a log page instead of a testimonials section.
Stack Python · SQL · FastAPI · Flask · Docker · Kubernetes · PostgreSQL · pgvector · FAISS · RAG · LangGraph · sentence-transformers · scikit-learn · TensorFlow/Keras · LLM evaluation · confidence calibration · Power BI · BPMN 2.0 · GDPR data lineage
See the CV Episodes Workflow blueprint How I build things The raw logs
Reads the small print in a privacy policy and flags the unfair parts. A live demo on HuggingFace, source on GitHub, and a 12-case regression suite that runs against the deployed service (not a notebook). When the verification fails, the page says so instead of pretending it ran.
Try it now → Case study → Weekly corpus of 9 EU + Big Tech policies → codeTwo bots that took over two jobs I used to do by hand every morning at Huawei. One reads complaints and decides which of five teams they belong to. The other runs the process. Below 0.55 confidence it hands the ticket to a human instead of forcing a label. 95.2% on the held-out set; the process was mapped in BPMN before any code.
Case study → codeA doctor asks in plain English and gets the answer out of that patient's own notes. The search filters to one patient before it ranks, so a different patient's note cannot surface. Self-hosted on the hospital's own machine, so the data never leaves.
Case study → codeThe moment an AI-generated model stops being the person you asked for, this catches it. Built so the agency can ship AI-generated imagery without the wrong face going out. A neural net reads the generated image against a locked reference, with a cutoff tuned on validation data only and 90 sealed test pairs.
Case study → codeTelecom mediation moves call records to fourteen downstream consumers. The constraint was never throughput, it was trust: a bad deploy means a customer is billed twice. On a 100k+ row validation pass, found a 14% reporting discrepancy nobody had seen.
On the CV →Everyone building on image models shows you the good outputs. So I ran the opposite test. Eighty garment images across four generation conditions, presented to human raters in a seeded shuffle, with condition and filenames masked and the rubric frozen before anyone rated anything. The rubric gates on one question first: is this image clear enough to judge at all. If the answer is no, the rest is skipped rather than guessed.
The interesting part is not the score. It is that an automatic OCR score and a human judgement disagree often enough to matter, and the only way to find out which one is right is to ask a person who cannot see the pipeline.
The instrument → Every call behind it →I make AI output verifiable before it ships. Every system here names the failure it defends against, and every number was measured, not estimated.