AI/ML engineer and founder, London. MSc AI and Medical Imaging at UCL, BSc Computer Science at City, University of London.
I investigate how models fail, build production AI systems and create products that people actually use.
kaszub.ski · LinkedIn · CV
| What it is | Evidence | |
|---|---|---|
| MIQA · code | MSc thesis: automated detection of geometric distortion in prostate diffusion-weighted MRI, and whether a model can tell when its own judgement is unreliable. PyTorch, patient-level cross-validation, deep ensembles, selective prediction. | 1,027 acquisitions from 627 patients; three-seed DenseNet-121 ensemble at mean AUROC 0.82 across five folds; a heterogeneous ensemble that recovers usable uncertainty, distilled to one model at a ninth of the inference cost. |
| AI Zuzi | AI social-media management and content platform, engineered for a client through NeuralTake. Django, React, PostgreSQL; seven third-party publishing integrations, payments, MFA. | About 50 business users. Security audit, encryption at rest, GDPR-aligned deletion and a zero-data-loss live migration from SQLite to PostgreSQL. |
| NeuralKite · live | Acquisition-target intelligence for fragmented UK private markets. Next.js, TypeScript, Prisma, PostgreSQL, Claude agents. | An email discovery and verification engine that grades every address into a risk tier with a written rationale, streamed live over SSE. |
| ChefBot | A cooking assistant built on GPT: photo of the fridge to recipe, dietary adaptation and a PDF guide, as one multi-step tool flow. | 25,000+ conversations and a 4.7 rating from 100+ ratings on the GPT store. |
Ongoing: an entry in the RSNA Knee Abnormality Detection challenge on Kaggle. Experiments in generative media and audience analytics are under kaszub.ski/experiments.
Fixes merged into quantitative-finance libraries, each found by testing the library against itself (Greeks against finite differences, parities, closed forms against Monte Carlo, one engine against another). Notes on each are on kaszub.ski/open-source.
| Library | Merged | What |
|---|---|---|
| FinancePy | 6 | American option pricer at zero dividend, date-schedule generation, forward-start option, Monte Carlo wrappers, variance-swap fair strike, FX error paths |
| vectorbt | 2 | Deflated Sharpe ratio (kurtosis convention and missing data), Numba/Rust engine parity |
| skfolio | under review | Parametric portfolio policies (Brandt, Santa-Clara and Valkanov 2009), online Schur allocation, rolling-window factor exposures |
| Riskfolio-Lib | under review | Risk-contribution finite differences for solver-based measures |
- NeuralTake Limited, Managing Director and AI Engineer, October 2024 to present. AI consultancy for professional services firms: workflow audits and fixed-fee automation builds delivered into the client's own stack.
- WorldQuant Brain, Quantitative Research Consultant, October 2023 to October 2024. Researched and backtested systematic trading signals in Python: risk-adjusted performance, turnover, decay and robustness.
Python, PyTorch, scikit-learn, pandas, NumPy, NiBabel · Django, PostgreSQL, TypeScript, React, Next.js, REST APIs · GCP, Docker, Playwright, CI/CD · English and Polish, both native.
Message me on LinkedIn. CV and references are available on request.