AI & machine learning
Biomedical data is small where it matters, enormous where it does not, and confounded almost everywhere. I build predictive models for that setting — and the tooling that makes working at this scale possible.
Two strands run through the work. The first is applied machine learning on biomedical data: proteomics, immune profiling, medical imaging and longitudinal clinical records. The second is engineering: LLM and agent systems, reproducible pipelines, and the infrastructure that turns a one-off analysis into something a team can rely on.
Strand one
Applied machine learning on biomedical data.
Everything below is work I have done rather than a list of methods I have read about. Most of it is published, or on its way there.
Plasma proteomics and immune-cell counts against incident disease in UK Biobank cohorts of several hundred thousand participants. Glycaemic phenotyping from HbA1c, medication and diagnostic codes; per-analyte models across thousands of proteins with Benjamini–Hochberg control; restricted cubic splines for non-linear dose–response; Cox models for time to event; and pathway enrichment to turn a protein list back into biology.
An AI framework built at HTW Dresden that predicts tumour spheroid status from radiomic features and, more usefully, predicts the point at which a spheroid escapes treatment control. The hard part was image-processing that stays stable while the underlying biology behaves strongly non-linearly.
Combining quantitative MRI features with simulated tumour–immune dynamics to predict time to relapse, and to identify which biopsy locations actually carry predictive information. Published in npj Systems Biology and Applications (2025): macrophage density at the tumour edge turned out to matter far more than where a biopsy is conventionally taken.
Where the biology is understood, write the mechanism down; where it is not, learn it. I combine agent-based and reaction–diffusion models with statistical learning, then use Bayesian inference and surrogate modelling to solve the inverse problems — recovering parameters that are not directly observable from data that is sparse and noisy.
Generating cohorts of virtual patients from a mechanistic model, then “measuring” them the way a clinic would — imaging volumes, proliferation indices, biopsies at chosen locations. It lets you ask which measurement is worth taking before anyone takes it, and it produces labelled data where real labels are scarce.
Global sensitivity analysis and uncertainty quantification on every model I ship: which parameters actually drive the prediction, how far the prediction can be trusted, and where the confidence interval is doing more work than the point estimate.
Strand two
Research software, LLM systems and automation.
Analysis is only half the job; the other half is the software that makes it repeatable. These are systems I have designed, shipped and run — for my own lab and my own research, which is the best test of whether the tooling is real.
Research software I developed for the Perakakis Lab at Uniklinikum Dresden, where I work: an application that digitises the group's in-house biobank data and makes analysing it substantially easier. Getting from the lab's own records to an answer used to be a manual job repeated for every new question — the application takes that procedure on instead.
A routing layer that lets each agent in a system pick its own model across providers — OpenAI, Anthropic and the You.com Research API — behind a single interface, with per-agent prompts, roles and fallbacks. Different tasks want different brains, and provider lock-in is a liability rather than a simplification.
Exposing domain-specific capabilities — literature scouting, peer-review critique, funding search — as MCP tools over stdio, so any MCP-capable client can call them. It turns a bespoke assistant into a set of composable tools that outlive the app they were written for.
A scheduled pipeline that queries PubMed daily, filters to a defined scope, generates structured summaries with an LLM — method, sample size, non-significant results and limitations, not just the abstract's headline — translates them, and publishes to a channel and a blog. Runs unattended on CI cron.
Parsers that decide encoding, delimiter, header row, decimal convention and column meaning from evidence inside the file rather than from configuration — because real-world exports lie, and the column that looks like the amount usually is not. The same discipline that keeps a clinical data dictionary honest.
Version-controlled pipelines with pinned environments, deterministic seeds, and provenance from raw extract to final figure — built so a colleague, a referee, or I in eighteen months can re-run the whole thing and get the same numbers. Large parameter sweeps run on HPC.
Principles
Four habits that decide whether a biomedical model is useful or merely impressive.
01
Most disagreements about a result turn out to be disagreements about who counts as a case. Thresholds, medication codes and diagnostic hierarchies get written down and defended first — the modelling is the easy half.
02
A feature that encodes the outcome, a split that crosses a patient, a preprocessing step fitted on everything — these produce the best-looking results you will ever get, and none of them generalise.
03
An AUC tells you how a model ranks people. A clinician needs to know whether the stated risk is the real risk. I report calibration and decision-relevant metrics alongside discrimination, or the number is not actionable.
04
Uncertainty quantification and global sensitivity analysis are not an appendix. They are how you find out whether a conclusion depends on the data or on a parameter someone guessed in 2011.
The useful question is rarely “can a model predict this?” It is “what would have to be true for this prediction to be wrong, and can we check?”
Get in touch
I work with researchers, clinicians, health-tech teams and bioinformaticians at the interface of biology, data science and personalised medicine — from a one-hour sanity check on a study design to a full modelling collaboration.
Details