Work
Markets work first, then software and machine learning. Each case gives the problem, what I built, how I tested it and its limits.
GB balancing cost forecasting
- Problem
- Forecast the total cost of the GB balancing mechanism for each half-hourly settlement period. Published GB work mostly targets imbalance price, so aggregate cost is a different target.
- What I built
- A Python pipeline over public Elexon BMRS and NESO data with lag, rolling and calendar features, comparing LEAR, XGBoost and LSTM under one evaluation scheme.
- How I tested it
- Rolling out-of-sample windows so no model sees the future, with Diebold-Mariano tests between models and against naive baselines. Forecast scores are not published before assessment.
- Limits
- Public half-hourly data only. The assessed working repository stays private. This page describes method, not results.
C++ order book simulator
- Problem
- Understand how a continuous limit order book matches and stores orders under price-time priority.
- What I built
- A single-instrument C++20 simulator for limit and market orders, partial fills, FIFO within a price level, cancellation by order ID, and read-only depth views.
- How I tested it
- Three CTest executables: written acceptance scenarios, randomised invariant checks, and seeded workload generator contracts. Debug build passed 3 of 3 on 29 September 2026.
- Limits
- A learning simulator. Benchmarks are single-threaded synthetic replay on one laptop. Not a broker connection, exchange or trading system.
gb-bm-data
- Problem
- GB balancing data is split across BMRS, Carbon Intensity and NESO. Some BMRS forecast endpoints ignore historical date parameters and only return live windows.
- What I built
- A typed Python client for system prices, demand outturn, generation mix and NESO datastore queries, extracted from the dissertation ingestion code so it can be reused alone.
- How I tested it
- A mocked HTTP test suite with no network calls. Confirmed live-only behaviour for DATL, FOU2T14D and demand/outturn/stream during real collection in August 2026; those routes raise LiveOnlyEndpointError.
- Limits
- Depends on public APIs. Not on PyPI yet; install from GitHub. Historical demand forecast recovery uses NESO, not BMRS v2.
GB energy forecasting reproductions
- Problem
- Check that published GB imbalance-price methods can be rebuilt from the papers before building a different target (aggregate BM cost).
- What I built
- Independent reproductions of Lucas (2020), Deng (2023) and Bunn/Ganesh on public Elexon and Carbon Intensity data, using gb-bm-data as the data layer.
- How I tested it
- Compared reproduced errors with the papers' reported figures. Only the Lucas MAE matched on the paper's own 2018 to 2019 window; other results are worse, and each README states why confidence differs.
- Limits
- Reproductions, not replications. Deng architecture and date range could not be confirmed with high confidence. Bunn/Ganesh only recovered 2 of 7 to 8 original variables from public sources.
Operational software for Goss & Parker Structural
- Problem
- Site-visit bookings, staff allocation, calendar events and reviewed client emails need to stay consistent across a small structural engineering firm.
- What I built
- As the firm's sole IT and software specialist, I independently built the booking and operations tools, including responsive client forms, validation, staff workflows, calendar synchronisation, duplicate checks, report generation and reviewed email drafts.
- How I tested it
- The scheduler repository contains 265 JavaScript test files, 44 recorded booking-form scenarios and a 7 of 7 live release gate. Company staff review workflows, test releases and help refine each system.
- Limits
- Client records, operational rules and source code remain private. The figures describe repository and release evidence, not customer-impact claims.
Reception system (prototype)
- Problem
- Small businesses miss calls. Aim: a phone handler that answers, works out why the caller rang and books an appointment on hardware the business already owns.
- What I built
- A prototype pipeline: Twilio for calls, Faster-Whisper for speech to text, spaCy for entities, a QLoRA fine-tuned local language model for the action, Pydantic validation, and Piper for speech back.
- How I tested it
- 111 automated pipeline checks covering bookings, cancellations and edge cases.
- Limits
- Prototype with no live client. Effect of quantisation on action quality needs more evaluation before any pilot.
UK home retrofit analysis
- Problem
- Estimate which dwellings have high retrofit headroom from open Energy Performance Certificate records, rather than only predicting the current A to G label.
- What I built
- Supervised models (Random Forest, XGBoost, SVM, logistic regression) on England and Wales EPC certificates, with a high-potential label based on current rating and score gap.
- How I tested it
- Temporal holdout rather than a random split. Positive class rate roughly halves from training to test, which a random split would have hidden. Nested cross-validation on the training window.
- Limits
- Labels inherit noise in EPC fields. Raw address-level files are not in the repository. Coursework project, not a policy deployment.
Retail sales KPI analysis
- Problem
- Turn a public online retail transaction set into validated KPI views.
- What I built
- Python ETL into MySQL, automated data-quality checks, KPI generation and a Streamlit dashboard.
- How I tested it
- Automated quality checks in the pipeline and unit tests in CI. Dashboard loads from the cleaned dataset after upload.
- Limits
- Coursework-style BI pipeline on a public Kaggle set. Not energy-market work.
CFRP defect detection (NCC challenge)
- Problem
- Segment carbon fibre composite micrographs and decide whether void severity fails manufacturing quality criteria used by the National Composites Centre challenge.
- What I built
- A U-Net segmentation pipeline with a source-stem split so augmented copies of the same image cannot leak into validation, plus a scoring layer that implements both NCC scoring scripts for comparison.
- How I tested it
- Confirmed 1,550 unique source stems against the 4,000-file dataset. Cross-checked scoring formulas against NCC's scripts on a synthetic case.
- Limits
- Hackathon analysis, not a production NDT system. Official team submission was separate from this personal working copy.