I build things I shouldn't be able to build alone.

Problem

I go to 100+ shows a year. There's no good way to find out if your friends are interested in the same ones — so I built one.

Built

A social events platform where you follow venues and friends, see what's happening, and make plans together. Auth, social feeds, event discovery, RSVP — the full stack. Then I rebuilt it from scratch: an LLM enrichment pipeline that scrapes raw event data nightly and uses GPT-4o-mini to categorize events, extract performers, and direct targeted Spotify and Knowledge Graph lookups. Claude Code enforces design consistency through a four-agent pipeline and ten specialized skills.

Learned

How to build auth and social feeds from zero. How much UX matters when real people use your app. And that rebuilding something from scratch — with a real design system and an LLM pipeline — teaches you more than building it the first time.

Problem

I was spending hundreds on Uber Eats, watching herbs go bad, and eating the same reheated meal three days in a row. Cooking something for 30 minutes and it not being tasty is a waste of time, money, and emotional bandwidth.

Built

A kitchen assistant that actually knows me. Through onboarding it learns my cooking days, cuisine preferences, skill level, and equipment. Full planning mode goes from pantry inventory to a week of meals — with prep lists and shopping lists generated automatically. Cooking mode for live sessions: timing, temperatures, substitutions, without touching persistent memory. Brainstorm mode for riffing on new ideas. A recipe importer that works across 400+ sites. An @ tagging system to pull in recipes and ingredients without bloating context. A 2,500-ingredient database with fuzzy and semantic matching for substitutions.

Learned

That the hardest part of a kitchen assistant isn't the AI — it's the data model. Ingredient normalization across thousands of recipes, inventory tracking that stays useful, and mode isolation so a quick 'what temp for the chicken?' doesn't derail your meal plan. This app is what made me realize the framework underneath was more abstract than a personal assistant — and that's what became Alfred.

Problem

I wanted my record player to show what's playing and generate visuals in real time. No reason other than it would be cool.

Built

A Raspberry Pi-powered audio fingerprinting and visualization system.

Learned

Debugging hardware with an AI pair programmer is a uniquely humbling experience. Also that the best projects start with "no reason other than it would be cool."

Problem

I've never had a personal brand. Never needed one. But I'd also never built a website, defined a design system, or orchestrated a multi-agent frontend build.

Built

The site you're looking at — dark mode, responsive, designed and shipped in a day using a four-agent Claude Code pipeline: design director, motion choreographer, page builder, and QA reviewer.

Learned

Design orchestration with AI agents, Playwright MCP for visual testing, devtools integration, and that defining a design system is the same muscle as scoping a product requirements doc.

Problem

I've played FPL for 10+ years. There's no shortage of data — 775 players, 38 gameweeks, hundreds of stats — but you can't just dump all of it into an LLM and hope for insight. I needed agents that could actually crunch numbers, not just summarize them.

Building

A BI layer on top of Alfred where agents act as data analysts. Instead of stuffing data into context, they write and execute Python in a sandbox — pandas for analysis, matplotlib for charts. Six domains unlocked: squad planning, player scouting, transfer market, mini-league rivalries, live gameweek tracking, and fixture analysis. I'd solved analysis — so I pushed further: if agents can crunch numbers, why can't they generate visualizations on demand? I spent years working alongside BI teams and wanted to prove that data viz on demand wasn't a pipe dream.

Will prove

Where the real line is between what LLMs are good at — reasoning, intent, formatting — and what they struggle with: math on 30,000 rows. Letting agents delegate computation to Python and keep interpretation for themselves is what makes the whole framework click.