Data engineer and analyst in London. I work with UK public data at the scale where the convenient approach stops working, and I publish what the data cannot support next to what it can.
31,430,611 rows · PostgreSQL
Every HM Land Registry price-paid record loaded in 8m11s. A 128 kB BRIN index beats a 211 MB B-tree on the same query, 1,687× smaller and 3.3× faster. The "new-build premium" turns out to be a flats composition effect. A script reproduces every figure.
18,080,573 rows per month · Python, Parquet
Checkpointed, resumable ETL: API paging into zstd Parquet, then into Postgres via COPY. No pandas anywhere, which is the point. Indexes are built after the load rather than during it, because that was timed rather than assumed.
215 providers · 23 months of Excel
The NHS four-hour A&E headline runs 14.3 points above what major A&E units achieve, and the gap widens exactly when performance is worst. I checked the mix-shift explanation and rejected it. Eighteen documented spreadsheet quirks, including a filename change that made the scraper return eleven months of twelve without raising an error.
9 ONS series · 664 months
Straight from the live ONS API. Vacancies per unemployed person are down 61% from the 2022 peak. The three headline rates have summed to over 100 every month since 1971, correlating with the unemployment rate at 0.988 — the cycle counted twice on two age bases, not demographic drift.
PostgreSQL · schema and access design
How a CRM holding several organisations' customer records is laid out: a tenant column leading every composite index, deduplication that stores the human's decision so the review queue can be told "no", and a reporting role that cannot reach credentials or money.
1,820,000 records · MSc dissertation, marked 75
VADER sentiment, NRCLex emotion, LDA topic modelling, Prophet forecasting and chi-square testing between communities. It ships a limitations document naming which outputs must not be quoted as rates.
Python · Whisper · OCR · FFmpeg · Docker
Decodes an Instagram reel into an edit template (cuts, motion, beat grid, timed captions), then re-renders the same edit with your own clips. Built with Arsalan Ahmed. Decode time cut from 599 to 290 seconds.
GitHub Actions · Cloudflare Worker · Gemini
Ranks the week's topics and drafts a carousel, then waits for a human tap on Telegram before publishing to Instagram and a LinkedIn company page. A guard stops any post going out twice. 71 tests.
PostgreSQL · Supabase · Android
A multi-tenant CRM for the agency's client work, plus CallTracker Pro, an Android app that logs sales calls into it, with web dashboards.
Playwright · n8n · Meta lead ads
WhatsApp Web outreach with image follow-ups, and a sync that moves Meta lead-form leads straight into the CRM.
Claude Code · MCP
A Claude content agent with a self-review quality gate, and my own agent desktop built on Pulse OS: terminals, voice, and chat over my notes.
I start from the question the person who has to act on the answer is actually asking, pull and clean the data in SQL and Python, run the analysis, and put the result where they will look at it.
Since October 2024 I have run client engagements end to end at Black Arrow Technologies. Three of those systems are live: a multi-tenant PostgreSQL CRM, a Python lead-scoring pipeline, and a workshop job-management system. I am also a director of Lateral View Solutions Ltd, a UK limited company.
I say what an analysis cannot support as readily as what it can. That is why the dissertation repository carries a written limitations section instead of a headline number, and why two claims in the Land Registry write-up now record that the original measurement was wrong.
Tools: Python (pandas, PyArrow, scikit-learn, PyTorch, NLTK) · SQL (PostgreSQL, MySQL, Supabase) · Power BI and DAX · Excel · R · SAS · Docker · Git