I turn research algorithms into production systems on AWS: Python, serverless and GIS. In my career I build the Python products behind real-time power-grid measurement. In my own time I apply the same engineering to conservation, building pipelines on open wildlife data that a field team could trust.

Every project here runs from committed or public data with no accounts or keys, is scored on data held out from tuning, and reports the result honestly, even when it is negative.

I'm open to projects in conservation and environmental data. You can reach me on LinkedIn.

camera-trap-mlops (in progress): decides, site by site, whether to trust an off-the-shelf camera-trap classifier (SpeciesNet, BioCLIP 2) or replace it with an adapted one. The test was fixed in advance, then ten camera sites held out from all tuning were scored once.

• A small classifier on DINOv2 embeddings beat BioCLIP 2 zero-shot at all ten sites, though by under 2 points at three. Balanced accuracy rose by 9.7 points on average (95% interval 5.5 to 15.0).
• It needs labelled photos where the off-the-shelf model needs none, and these sites share a study region, so the result says nothing about a new region.
• Still to build: a per-site trust report and label-free signs that the decision needs revisiting.

collar-telemetry: an alerting pipeline for wildlife GPS collars. It turns late, duplicated and missing fixes into alerts for possible deaths, silent collars, unusual movement and park-boundary crossings, scored against real recorded deaths in public Movebank data.

• Every held-out vulture death raised an alert (10 of 10), but only 3% of fast mortality alerts were real (95% interval 0.9-8.0%). The slower "potential death" alert did better: 29% (8.6-83%) with ideal data arrival and 43% (15-88%) with daily satellite uploads. With so few deaths, those intervals are wide.
• On Kruger elephants, every "left the park" alert came from the unfenced western edge shared with private reserves: the boundary matters as much as the detector.
• Deployed to AWS with Terraform (SQS FIFO, Lambda, DynamoDB, SNS), replayed over 3,466 animal-months with every incident matching the local run, then torn down. The full run cost $0.23; the idle queue trigger, not the animals, is most of a small project's bill.

vessel-anomaly: flags vessel behaviour worth a fisheries analyst's attention in public AIS tracks from the US Northeast: fishing in closures that ban the gear in use, unexplained transmission gaps, encounters at sea and loitering. Thresholds were frozen before a held-out block of 542 unseen vessels was scored once.

• A coverage mask records where receivers cannot see; it removes two-thirds to three-quarters of gap alerts that would otherwise be reception holes, not vessels going dark.
• Gaps planted into real tracks were found 71% of the time, and a classifier shows planted gaps are easier than real ones, so that is an upper estimate. The other detectors caught every planted event and none of the planted near misses.
• Runs as a restartable streaming service in Docker that matches the batch run event for event, even after a mid-run kill. Costed from measured work against AWS prices: batching costs about $0.76 a month at regional volume, against $42 for processing each message on its own.

• population-model: how far can noisy aerial counts of Kruger herbivores be trusted? Two forecasting models failed to beat naive baselines, both documented as negative results; a third version shows survey precision matters more than frequency for catching a real decline in roan, sable and tsessebe.

Python · AWS (Lambda, SQS, DynamoDB, SNS, EventBridge) · Terraform · Docker · polars and DuckDB · GIS (shapely, GeoJSON) · state-space models and Kalman filters · pytest · GitHub Actions

MEng Computer Science (Distinction), University of Strathclyde. Certificates in GIS and in ecology and ecosystem conservation; ZSL Academy participant.