Things I’ve built, and why

GB Grid Scenario Tool

I was drawn to reinforcement learning through the demos of agents learning to drive cars and play games. Weather and climate systems do not map cleanly to the same RL framing. Energy grids, where weather is a core input, were a much better fit. So I started building an RL environment for GB electricity dispatch. The scope grew quickly. What began as ‘can an agent learn merit order?’ turned into a serious exercise in data assimilation from public sources, and a test of how far you can get with human-led architecture and AI execution. The question that drove the whole project was simple: how much real skill can you get from only public data?

What I ended up with is a browser-based DC power flow model of the GB transmission network, built from NESO open data. It runs at two resolutions: 27 TNUoS zones, and 82 FLOP zones that match NESO’s internal boundary analysis resolution. Three dispatch modes show the progression from naive to realistic. A simple all-on dispatch. A cost-merit-order dispatch with minimum stable level constraints. And a full LOPF that minimises system cost subject to boundary flow limits, solved using HiGHS compiled to WebAssembly. Weather scenarios are driven by 34 years of ERA5 climatology, and you can drag a slider from 2024 to 2035 to see planned network reinforcements appear. The B6F boundary flow validates within 2% of NESO’s published ETYS transfer capabilities at the 27-zone resolution.

The RL layer sits on top of that. I trained PPO agents on nine years of real GB generation and demand data, plus Elexon wholesale prices and ERA5 weather, with the grid model as the environment. I compared MLP, CNN, and 27-zone zonal architectures, the largest with 448 thousand parameters reading 72 ERA5 weather channels. Iterative reward design moved agents from spending around £30 billion a year on dispatch to operating within 6% of real grid costs. The interesting finding was that spatial CNN observations could not improve on MLPs without topology-aware action spaces. The action space was the bottleneck, not the observation space.

• JavaScript
• React
• Leaflet
• WebAssembly (HiGHS)
• Python
• Stable-Baselines3
• SLURM
• ERA5

Read the full build story · Live demo · Source on GitHub

Quantum ML for atmospheric regression

Quantum computing as a research direction is interesting precisely because nobody yet knows what it will be good for. I wanted to test whether variational quantum circuits could do anything meaningful on real atmospheric data, or whether they were a solution chasing problems. I built VQC regression models on ERA5 temperature fields and on the Lorenz ’63 system, using PennyLane, and ran depth sweeps, noise ablations, and cross-validation on the Lorenz attractor.

The honest answer on raw performance: VQCs did not beat MLPs of comparable size on either problem. The classical baseline was strictly better. Where the project landed instead was on interpretability. Per-qubit structured VQC regression admits a Fourier decomposition that lets you read off, qubit by qubit, which frequency components a given circuit is using. That turned out to be the more useful contribution. Not a faster predictor, but a sharper lens on what these circuits are actually doing when they fit data. I documented the single-seed limitations of the experiments and the VQC versus MLP gap explicitly in the handoff notes.

• Python
• PennyLane
• ERA5
• Lorenz ’63

Climate Data Quickstart

Climate data is fragmented. Every dataset has its own access pattern: a CDS API for ERA5, an ESGF queue for CMIP6, raw FTP for HadCET, a separate Earthdata account for NASA products, and so on. Getting a working environment for any one of them is solvable in an afternoon. Getting working environments for all of them, with credentials configured and credentials documented and a stable place to put the files, is a quietly large amount of plumbing. For most early-career researchers, that plumbing is what stands between them and the actual science.

Climate Data Quickstart is a local desktop app and script library covering 19 datasets across five categories. Reanalysis (ERA5 variants and ARCO-ERA5), climate projections (CMIP6 via CDS and ESGF), observational temperature (HadCET, HadCRUT5), precipitation and station observations (GHCNd, CHIRPS, E-OBS), and specialised products like UKCP18, GloFAS, GPWv4, and ECMWF Open Data. It bundles setup scripts for Windows, macOS, Linux, and conda, exposes a Streamlit interface for downloading and exploring data, and supports lazy loading via Earth Data Hub for the larger products. It was built using a three-stage agentic pipeline: dataset schema extraction, code generation, then validation against each provider’s actual API.

• Python
• Streamlit
• xarray
• cdsapi
• NetCDF/GRIB/Zarr

Source on GitHub

City Climate Stripes

Ed Hawkins’ warming stripes are one of the most effective single climate visualisations ever made. I wanted to extend the idea to individual cities, with a few extras: switching between annual and seasonal anomalies (DJF, MAM, JJA, SON), toggling bars versus stripes, adjustable baseline periods (1850 to 1900, or 1961 to 2010), and a fixed-versus-auto colour scaling. Built on Berkeley Earth gridded temperatures with GeoNames city coordinates, runs entirely client-side, exports PNGs. A small exploratory thing, not a scientific dataset.

• HTML
• JavaScript
• D3
• Berkeley Earth
• GeoNames

Read more · Live demo · Source on GitHub

Climate Playbook

Climate Playbook is an interactive climate education platform aimed at UK schools, Key Stage 1 through Key Stage 5. The motivation is straightforward: most climate education at school level is either too patronising for the older students or too jargon-heavy for the younger ones, and very little of it puts the physics in the children’s hands. I want a set of modular, MDX-based lessons that scale with the reader, with interactive figures showing the actual mechanisms.

This is a slow project. Pedagogy is hard, and designing content that works across a six-year age range alongside an MSc workload is realistically a multi-year arc. I’m framing it honestly as a work in progress. British English throughout, no edutainment, no gamification for its own sake.

• Astro
• MDX
• D3
• vanilla JavaScript

Electricity demand forecasting (India)

An MSc module sub-project. State-level electricity demand prediction for India using ERA5 weather variables, with linear regression, XGBoost, and LSTM models as a progression in capacity. I used temporal cross-validation throughout and mapped per-state R-squared as a choropleth, which made the geography of the problem obvious. Weather-driven prediction works well in the southern states where demand is dominated by cooling load, and badly in the industrial north where demand is dominated by non-weather factors. A short exercise in building the whole pipeline end to end on a non-UK grid.

• Python
• scikit-learn
• XGBoost
• PyTorch
• ERA5