I build and evaluate custom ML and LLM systems for sensitive-data settings. My recent independent research explores medical imaging and rare-disease AI. I take projects from feasibility studies to production-oriented prototypes, with privacy and real-world constraints considered from day one.
I take on freelance ML and LLM engineering, from feasibility work to production-oriented prototypes. Current independent research applies these methods to medical data; the same engineering also serves other sensitive and regulated settings.
Train and adapt models — from vision encoders to language models — around your task and data. Build a reliable baseline, compare approaches fairly, and carry promising results into a working prototype.
Build and evaluate language-model workflows for document-heavy tasks, including extraction, retrieval and custom small-model approaches. Data protection and the limits of the source material stay explicit.
De-risk a new approach before a larger investment: establish a reproducible baseline, test it against the right cases, and turn promising results into maintainable prototypes. My current independent research applies this approach to medical imaging and rare-disease AI.
The technical foundation behind the services: hands-on studies with real data, fixed seeds and reproducible setups — including what didn't work. Skim the takeaways, or read the full depth.
A practical investigation of structured findings, unstructured text, images, provenance bias, and the clinical meaning of a concrete Moyamoya case study.
A frozen, self-supervised foundation model — trained only on everyday photos, never on medical images — detects intracranial hemorrhage on head CT at ~0.94 AUC. What drives the transfer isn't what I bet on.
LeCun's video recipe makes patches self-organise into structure for free — and trains at a fraction of the compute. Ported to text, the efficiency transfers but the structure collapses: every token onto one gist. Across an ablation, a word-salad test and 14 datasets, one measurable data property predicts whether the recipe grows token structure at all — before training.
I combine more than 16 years of software and ML engineering, starting as a working student developer in 2010, with a PhD in deep reinforcement learning (2× IJCAI). Alongside freelance engineering, my recent independent research examines medical imaging and rare-disease AI using public data.
My engineering work is not limited to healthcare: the same methods apply to other regulated and privacy-sensitive settings, including finance, legal and pharma.
I go all-in on a clearly scoped problem — defined goals, a defined outcome. You get full focus and a clear endpoint, rather than open-ended hours. Remote, with on-site where it helps.
I'd rather be the engineer you come back to again and again than one permanently parked on-site. A loyal, long-term relationship — built from many focused engagements.
Straight answers on what's feasible, the simplest thing that works, and no overselling. The data decides.
For recurring needs I can build, deploy and run a model as a service — you use it and pay for what you use, while I keep improving it.
And we will find out if AI can be a solution. Let's talk.