~projects

A personal weather station that got out of hand. Five services doing forecasting, air quality, camera timelapses and dashboards, running on a Raspberry Pi at the house.

Turns support tickets and a knowledge base into training data for a Tier 0 LLM support chatbot. Built for real use at DeepMind first.

An engineering pipeline that runs academic qualitative content analysis at scale, on Hansard, Parliament data, and AI incident reports.

Criminological analysis of street-level crime data for any UK postcode, using data.police.uk, postcodes.io and deprivation scores. Ties the OU Criminology degree directly to the technical work.

A research journal in Jupyter notebooks, working through famous maths problems (Collatz, Hadwiger, Goldbach, Waring and others) not to solve them, but to understand exactly why each one resists proof. One of the favourites.

Benchmarks the self-play noughts-and-crosses agent against DeepMind's own OpenSpiel MCTS bot. 199 draws out of 200 games, and about 1,600x faster per move. A ten-years-later answer to a DeepMind technical test.

Writing down what got absorbed by osmosis over 13 years at DeepMind, before it stays tacit knowledge. Notebooks so far: LSTMs, SGD, Transformers, Attention, GANs, RLHF, Actor-Critic, Markov Processes, MCMC, KNN, Bagging/Random Forests, Tokenisation.

Gathers Linux system evidence against six Annex A controls and chains it into a tamper-evident ledger, so an auditor gets cryptographic proof nothing was touched after collection.

A hash-chaining and Merkle tree library, the tamper-evident core the ISO 27001 tool runs on. Deliberately excludes the consensus and mining machinery nobody needs.

A packaged Python tool for analysing GEDCOM family tree files as directed graphs. Stats, validation, relationship-finding, source-checking. Parked deliberately: no roadmap, it grows only when family tree research hits a wall it can solve.

Genealogy project that pairs with gedcom-toolkit. Still needs a better name than this one.

Collecting TfL delay data since June 2026, over a million records now. The dataset's big enough to start building predictive models on top of the three analysis notebooks already done.

Two-way sync between HubSpot Deals and Airtable that doesn't fall into the sync-loop trap. Content-hash state store, conflict resolution, full test suite.

A bare-metal, multi-architecture Kubernetes cluster for keeping MLOps skills sharp without a cloud bill. K3s, MinIO, MLflow, GPU training.

A noughts-and-crosses AI trained with TD-learning from scratch, no libraries. Published as a LinkedIn article. See spiel-check for how it holds up against DeepMind's own OpenSpiel.

A small dockerized time service. Dockerfile, tests, GitHub Actions, nothing fancy, just needed one that worked.

DfE dataset analysis of children-in-need and looked-after-children pathways. A notable Shropshire outlier.

The same DfE pathway analysis, applied to special educational needs data.

Watches a local ADS-B feed and alerts on rare aircraft types. Started because the Goodyear Blimp flew over the house. Warm-up baseline is over and the dataset's now big enough to start building rarity predictions on top of the alerting.

Home for various Home Assistant scripts, including an audit tool for migrating between HA servers without a bloated restore from backup.

Feeds Home Assistant telemetry into InfluxDB and exposes it over MCP. Written, not yet deployed.

Merges log streams from multiple sources, local files and Docker logs, into one chronological, human-readable timeline.

Reads logtimeline's output and explains what the log events actually mean, in plain English. Waiting on logtimeline to ship first.

A home status board on an e-ink display: occupancy, appliances, weather, drying times, updating every five minutes. Paused while Home Assistant migrates to a new server.

MicroPython on a Pimoroni Galactic Unicorn LED matrix, scrolling Home Assistant notifications with colour-coded urgency.

A decision-logging tool, mostly an excuse to prove out a GCP / Cloud Run deployment pipeline. Stalled at the pipeline stage.

A client for the Capacities note-taking API. Paused while the API was in beta, worth revisiting now it's hit a stable 2.0.

Pulling planning and company filings on a local development. Stuck on scanned PDFs that need OCR.