none of this
was necessary._

I wanted to actually see where my money goes, so I built this. It pulls every transaction from my Up Bank account, categorises them, and renders the whole picture through custom D3 charts.

Go backend, PostgreSQL, vanilla JS frontend. Sankey flow chart for spending by category with month-on-month trend deltas, a streamgraph for net worth over time where you can click into individual bands, savings goals with progress bars, and a savings rate sparkline in the KPI strip. Categorisation uses pattern matching first, then an LLM for anything ambiguous — a Claude agent runs at 3am on a Mac Mini to classify stragglers automatically.

730+ tests across 14 packages. 3,800+ transactions categorised across 3 years of real spending data. Not a side project I abandoned — I use it every day.

Started as a Gmail poller that logged my bills. Kept going. Now it's a personal AI assistant that runs on a Mac Mini under my desk.

It polls Gmail via IMAP, classifies emails with a local Ollama model, and routes them — bills go to spending-app, interview invites go to Google Calendar, job offers ping me on Discord. A nightly reclassifier sweeps up anything the first pass missed, using Claude as the escalation brain.

Also runs live chat bots on Discord and Instagram with a RAG pipeline over my personal knowledge base, and serves a FastAPI booking API that powers the date page on this site. 930 tests. Nine hundred and thirty.

Everything I build ends up running on a single NUC sitting in my room. Four public sites and ten-plus containers. Caddy handles routing, Docker Compose runs each service, and GitHub Actions deploys everything automatically when I push.

I wrote a Go config generator so adding a new site is just a YAML file and a PR — it produces the Caddyfile, compose config, and security hardening (cap_drop ALL, no-new-privileges, resource limits, log rotation) automatically. Deploys build the image, push to GHCR, trigger the NUC to pull and swap the container, and roll back if something breaks.

PostgreSQL backups run daily with offsite R2 sync. WAL archiving enables point-in-time recovery. Autoheal restarts anything that goes down. Beszel watches resource usage. Uptime Kuma monitors everything from the outside. This site is running on it right now.

My mate Norton wanted to build a florist discovery site for his girlfriend. I said I'd help. Ended up doing the architecture, security hardening, deploy pipeline, and hosting it on my NUC.

Next.js, TypeScript, Supabase, Tailwind. Search page with filters, florist profiles, enquiry form, analytics dashboard, and a bouquet picker quiz with weighted scoring across eight bouquet types. I went through and hardened the API routes, added a Valkey SWR cache, image optimisation with Sharp, and made sure Supabase wasn't leaking internals to the client. 267 tests.

His project, his repo. I just made sure it wouldn't fall over and added a few hundred tests. It will still fall over. Norton must really love his girlfriend.

First actual client project. DeltaLilac is an artist who needed a proper home for their work, something more than a Twitter feed or a Linktree.

Full artwork database with a tag-filtered gallery, pagination, featured showcase, commission status toggle, and light/dark mode. Next.js, PostgreSQL, Tailwind. Images served from Cloudflare R2 CDN with AVIF/WebP optimisation and blur-up placeholders. Admin panel with drag-and-drop upload and inline editing, hardened with HttpOnly cookie auth and Postgres-backed rate limiting.

The gallery works, the dark mode is suspiciously good, and the admin panel is more secure than it needs to be.

Automated job application pipeline across seven sources — LinkedIn, Seek, Indeed, Glassdoor, plus direct ATS scraping via Greenhouse, Lever, and SmartRecruiters for companies like Canva, SafetyCulture, and Immutable. Each listing lands in a SQLite store, gets auto-scored by Claude against a target profile, then passes through a tailoring pipeline that generates a bespoke resume and cover letter for each posting.

Python throughout. WeasyPrint renders the final PDFs. Playwright handles browser automation for submission. A FastMCP server exposes the pipeline to Claude Desktop for manual intervention. Dockerised dashboard running on the NUC. Daily cron with a dry-run gate to prevent accidental mass-applies. 163 tests.

Private repo — it knows too much about my job hunt to be public.

I kept getting asked for feedback forms by people who didn't want to sign up for Typeform. So I built one.

Go backend, PostgreSQL, single-page HTML frontend embedded in the binary. Anyone can drop a name, context, written feedback, and a star rating — no account needed. Submissions persist to Postgres and trigger an email notification. Bot protection via Cloudflare Turnstile, CSRF prevention with HMAC-SHA256 double-submit cookies, per-IP rate limiting at 5 submissions per hour.

The feedback form for this site runs on it. Nobody has abused it yet, which means either the rate limiting works or nobody has feedback.