Machine learning engineer working on computer vision, explainable AI, and production forecasting systems. First-Class Physics, then an MSc in Data Science & AI (Distinction). Focused on what breaks after a model ships: temporal data leakage, silent target drift, and CI gates that catch degradation before merge.
A containerised forecasting service for Great Britain's half-hourly grid carbon intensity: daily 48-hour forecasts, an automated CI accuracy gate, and a public scoreboard benchmarked against the National Energy System Operator (NESO).
- 18 consecutive days recorded (
2026-11-02to2026-11-19):lgbm-v2scores27.06gCO2/kWh macro MAE (n=278settled periods) vs.38.64forpersistence-v0(n=1,622), while publishing the frozen benchmark beside NESO (39.118vs.11.666, vintage unverified). See the live scoreboard. - Leakage tests with mutation checks: Features are restricted to values available at the
00:00 Europe/Londonissue instant with a 3-hour actuals lag, enforced by tests that fail when future rows are injected. - Proven CI accuracy gate: PR #1 submits a degraded feature set (
48.682>43.3379gCO2/kWh threshold) where lint, typing, 337 unit tests, and Docker smoke checks pass, and CI fails specifically onmake eval.
gridcast repository · Live scoreboard · Failing-gate PR #1
| Project | What it is | Stack |
|---|---|---|
| gridcast | 48-hour GB carbon-intensity forecast API: CI accuracy gate, nightly dual-model scoreboard, non-root Docker | Python, FastAPI, LightGBM, Docker, GitHub Actions |
| Sports-Analytics | End-to-end CV pipeline over match footage: detection, tracking, pitch-line segmentation, jersey OCR, and a classical fallback for cross-sport domain shift | PyTorch, YOLOv11, ByteTrack, OpenCV, EasyOCR |
| Cryptogame | Cross-platform cryptogram puzzle app: 12 subject archives, daily cipher, timed mode, coin economy, 24 test suites | React Native, TypeScript (strict), Jest |
| Cricket outcome modelling (MSc dissertation) | XGBoost, Random Forest, SVM, and Swin Transformer over ~38,000 deliveries with SHAP attribution | scikit-learn, XGBoost, PyTorch, SHAP |
| Cricket sequence modelling (research, paper in prep) | LSTM, BiLSTM, GRU, and CNN with attention pooling across 5/10/15-ball contexts against gradient-boosted baselines | PyTorch |
| Medical image classification | ResNet18/50, DenseNet121, and Swin benchmark on a 7-class white blood cell task | PyTorch |
| FairRide | On-chain ridesharing: real-time reverse auction in Solidity on 300ms blocks, built in seven hours at Monad Blitz London | Solidity, Monad |
- Leakage checks before models: Forecasting and evaluation pipelines start with a failing test for temporal or split leakage before features or training code land.
- Guards verified by sabotage: Every test guard or CI gate gets a negative control so passing for the wrong reason fails loudly.
- Immutable evaluation fixtures: Benchmarks pin their ground-truth split and artifact SHA-256 hashes so daily archive updates cannot move published numbers silently.
Python / PyTorch / OpenCV / YOLO / scikit-learn / XGBoost / LightGBM / SHAP /
FastAPI / Docker / GitHub Actions / SQL (PostgreSQL, MySQL) / MongoDB /
AWS (SageMaker, S3) / Power BI / TypeScript
Machine Learning Engineer, Data Scientist, and Computer Vision Engineer roles in London. Full right to work in the UK (no sponsorship required).