Problem 01
You have a cohort and no idea where the signal is
Thousands of analytes, hundreds of thousands of rows, a dozen plausible endpoints, and
an analysis that keeps producing findings that evaporate on the next dataset.
- Approach
- Nail the phenotype and endpoint definitions first. Screen across all analytes with proper multiple-testing control, model non-linear dose–response rather than assuming linearity, then confirm with time-to-event models and pathway-level interpretation.
- Outcome
- A ranked, FDR-controlled biomarker panel with effect sizes and confidence intervals, the biology it implicates, and an honest account of what the cohort cannot tell you.
Problem 02
Your model looks excellent in the paper and fails in the clinic
High AUC on the development set, disappointing performance everywhere else. Usually this
is leakage, distribution shift, or a model that ranks well but whose probabilities mean
nothing.
- Approach
- A structured audit: data-flow review for leakage, patient-level split integrity, recalibration, external or temporal validation, subgroup performance, and decision-curve analysis at the thresholds you actually use.
- Outcome
- A validation report naming the specific failure modes, a recalibrated model where recalibration is enough, and a clear statement of the population the model is licensed for.
Problem 03
You need to know whether the study is worth running
An expensive trial, a costly assay, or a new measurement — and no way to know in
advance which arm, endpoint or sampling schedule gives you a chance of detecting anything.
- Approach
- Build a mechanistic model of the process, generate a virtual cohort, and run the study in silico under realistic noise and dropout. Vary the design; see which versions can actually resolve the effect.
- Outcome
- A design recommendation backed by simulation — sample size, timing, which measurement carries the information, and which planned measurements are not worth their cost.
Problem 04
You have imaging and outcomes but no pipeline
MRI, histology or microscopy sitting in a folder, an outcome column, and a cohort far too
small for the deep-learning approach everyone suggests.
- Approach
- Reproducible segmentation and quantitative feature extraction, feature selection that is stable under resampling, and models sized to the cohort you have rather than the one you wish you had.
- Outcome
- A running image-to-prediction pipeline, the stable feature set, and validation that survives the small-n regime.
Problem 05
Your analysis cannot be re-run
The results exist, the person who produced them has moved on, and nobody can reproduce
the figure — which becomes urgent exactly when a regulator, a referee or an investor asks.
- Approach
- Rebuild the analysis as a version-controlled, environment-pinned pipeline with provenance from raw extract to final figure, documented decisions, and tests on the steps that matter.
- Outcome
- A pipeline anyone on the team can re-run to identical numbers, plus documentation of every judgement call the original analysis made silently.
Problem 06
Your data exists, but not in a form anything can analyse
A biobank, a registry or a cohort whose records were never collected with analysis in
mind, so every new question begins by assembling the data by hand.
- Approach
- Design a data model that fits how the data is actually collected, build the path that digitises it, and put the analysis behind an application the team runs itself. I built this for the Perakakis Lab's in-house biobank.
- Outcome
- A structured dataset and the software that maintains it — plus analyses the team can run directly, instead of rebuilding them by hand each time.
Problem 07
You want to use LLMs without inheriting their failure modes
Everyone wants AI in the product. Nobody wants a confident, fluent, wrong answer attached
to a clinical or scientific claim.
- Approach
- Scope the task to something checkable, ground it in retrieval over sources you control, structure the output so it can be validated automatically, and build an evaluation set before shipping. Multi-provider routing so one vendor's outage or price change is not an outage of yours.
- Outcome
- A working pipeline with an evaluation harness — measured accuracy on your own task, cited sources, and a clear boundary around what the system is allowed to assert.