Selected work
A product in progress and four technical case studies, each separating the question, method, reported result and what the evidence cannot prove.
My evaluation lens
I use the same five questions to connect technical quality with what a person can safely decide.
Verdict: research prototype, not a clinical product.
Inspect method and code ↓
An evidence-first B2B SaaS designed to help UK fire-safety contractors identify relevant public procurement opportunities and likely contract-renewal windows.
Published opportunities are never relabelled from a renewal forecast. Evidence remains attached to the signal.
Profiles, deterministic matching and renewal intelligence focus the product on UK fire-safety procurement.
The local build and browser journey have been validated; hosted CI, target VPS acceptance and production readiness are not claimed.
AI × Lifestyle · Movement Lab
An interactive experiment with a pretrained pose model, local image processing and visible confidence limits.
Try the demo →01 · RAG and evaluation
A research retrieval-and-answering prototype in which every response stays connected to evidence and a specific patient in a synthetic dataset.
Read the full case study →Build an auditable RAG workflow for NHS-style records without using real patient data.
TF-IDF retrieval, qrels, citation-aware answers, safe responses when evidence is insufficient, pytest checks and documented assumptions.
Reported retrieval results: Precision@3 0.583, Recall@3 1.000 and MRR 1.000 within the project’s defined evaluation.
The dataset is synthetic and the system is an academic prototype, not a clinical product or medical device.
02 · Computer vision
A NumPy implementation of perceptron and feed-forward networks focused on label quality, error analysis and model comparison.
Read the full case study →Understand the mechanics of neural-network learning by implementing the essential parts without a high-level framework.
Backpropagation, regularisation, early stopping, linear-versus-nonlinear comparison and checks for inconsistent labels.
Primary accuracy on the original held-out labels: 0.4584. Separate aligned-label diagnostic: 0.8862.
The score belongs to the Sign Language MNIST benchmark; it is not the performance of a real-world camera-based sign-language system.
03 · Machine learning
A reproducible regression and classification pipeline with baselines, feature engineering, validation and automated tests.
Read the full case study →Separate predictive signal from an appealing but weakly generalising story about music data.
CatBoost regression, Linear SVC classification, model comparisons, diagnostic figures and data-quality tests.
Reported benchmark RMSE of 6.52657 and classification accuracy of 0.57142.
Results depend on the dataset, split and target scale. The work documents overfitting risk and is not a production recommender.
04 · Optimisation
A chronological evaluation of a 15-parameter strategy with transaction costs and several comparison baselines.
Reduce the risk of fitting a strategy to history and evaluate it on unseen, later data.
Time-respecting train/test split, random-search, fixed-rule and buy-and-hold baselines, plus explicit transaction costs.
Reported out-of-sample return of 7.50% with maximum drawdown of 9.99%.
A research project, not investment advice. Historical performance does not predict future returns and needs its period and benchmark for full interpretation.