An AI hiring platform still ran hiring as a fully manual process, done by the founder.

Built the full AI agent layer in Python on the Anthropic Claude API: around a dozen LLM agents and an orchestrator, with 1,100+ unit tests. Added automated bias audits and evidence checks so every AI decision traces back to its source material.

The founder-run process became an automated agent system, hardened over 3 rounds of production test runs, with an evaluation harness that blocks deployment when quality drops below baseline.

Engineering knowledge was spread across thousands of technical manuals that reference, replace and build on each other, so one answer could sit across many documents.

Led the project. Built a pipeline that extracts and cleans text from the PDFs, a custom spaCy model that spots document codes, and relationship extraction that labels whether one manual references or supersedes another. Then a RAG assistant on top.

Thousands of linked manuals became one knowledge graph engineers can query, supporting infrastructure asset management.

Most ML work is a config file pointed at someone else's library. I wanted to prove I understood the machinery underneath.

Wrote CUDA kernels by hand (tiled matmul, warp-shuffle softmax, tiled attention) and my own autograd engine, then trained one decoder on text and on MovieLens 1M without changing it.

Tiled attention runs 25.7x faster on a Tesla T4 than my CPU version at sequence length 1024, it scores 98% on MNIST, and it runs live as a movie recommender.

Rain, fog and snow made roadside CCTV hard to read, which weakened road monitoring.

Trained CNN models on 10,000 CCTV images to recognise the weather in each frame, and a model to remove it from the picture.

98% weather classification accuracy, making road monitoring more reliable.

Crash investigators need to know exactly which car was involved, and how safe it is, often from a single photo.

Built a two-branch ResNet50 that predicts make, model and year with calibrated confidence, trained on Stanford Cars (8,144 images, 196 classes), and matched to Euro NCAP safety ratings.

46 ms model inference per image. Make 96.4%, model 91.0%, year 89.1%, and 81.7% with all three right.

£10,000 of discounted shoebox gifts had to reach 50 community groups, and every order went through a spreadsheet.

Designed and built the whole system solo: an online order form, automatic payment and confirmation, and a live stock dashboard with low-stock alerts.

£10k of products and 50 groups handled in one place, replacing the spreadsheet to cut admin time and order errors.