Self-taught, working in Python. Data pipelines, PostgreSQL/PostGIS and geospatial data. Heading for data engineering.
sentinel2-brownfield-stoke is the one to look at — a Sentinel-2 satellite pipeline built to find unregistered brownfield land in Stoke-on-Trent. Copernicus API ingest with token refresh, cloud masking, coordinate reference transforms, spectral indices, connected-component clustering, PostGIS storage with a migration-driven schema, 400+ tests, and CI running against a live PostGIS container.
Live demo — deploys from a bare checkout with no database connection and no secrets. It sleeps when idle, so give it a few seconds to wake.
It doesn't work, and the repository explains why. Registered brownfield land turns out to be vegetated; the detector was looking for bare ground; zero of 352 known sites pass its threshold. Establishing that meant retracting recall figures I'd been reporting for three months, tracing the cause to a comparison I'd never made between two numbers recorded in the same notebook, and correcting the record across six notebooks and the README. I'd rather show that than a screenshot of something that runs.
I stopped there rather than scaling it. The supervised route was closed on an effect size of about 0.03 SD between matched and unmatched candidates — there was no label to learn — and a UK-wide version wasn't worth building on a method that fails at one council.
ftse-portfolio-rebalancer — ridge-regularised rebalancing under FCA position, sector and liquidity limits; κ reduced from 2,862 to 62.7. An external review found a one-character error that invalidated the headline result while all 21 tests passed. Diagnosed, fixed, re-verified, and I write the failing test first now.
tfl-optimiser — a linear programme allocating 15 buses across 25 routes covering 50 stations, solved with HiGHS. Uses LP relaxation rather than integer programming, which is a documented trade-off rather than an oversight.
week3-linear-solvers and week3-gaussian-implementation — a deliberate pair. One picks its own strategy from matrix shape and condition number; the other implements the same mathematics by hand and benchmarks it against LAPACK. NumPy wins, which is the argument for using NumPy.
Also here: Markowitz portfolio optimisation under no-shorting constraints, NHS resource-allocation models, and two small vector CLI tools from the start of all this.
Python · SQL · PostgreSQL/PostGIS · Docker and docker-compose · GitHub Actions · pytest · NumPy, SciPy, pandas, scikit-learn, matplotlib · rasterio, pyproj · Streamlit, Folium
Everything above is self-directed project work from 2026. My degree is in History, not computer science, and I'd rather you read the code than take my word for any of it.
Open to remote or hybrid work — based near Newcastle-under-Lyme, Staffordshire. Data engineering is what I'm aiming at; also interested in data science and geospatial roles.