What I'm up to /now.

An AI trading experiment — Claude makes the buy / sell / hold calls, a small web app handles the data and execution, and a paper-trading account at Alpaca is the live target (no real money yet). 24/30 paper trades executed (as of 2026-05-14); 6 more needed to gate the Phase G live switch. Only Test 4 ($10k portfolio) is active; three $100 test portfolios deactivated (spent). Fixed: BTC/ETH removed from Claude prompts (StockHistoricalDataClient returned wrong $35 price for crypto tickers). Decision modes, oversight activity, pause/resume UI, sign-out all shipped.

The compliance platform I wished existed when I was importing — one place to manage CPSIA certificates, Prop 65 labels, PFAS tracking across 15 states, and FSVP audits for food brands. All four compliance modules are feature-complete: CPC version history + supersession, expanded PFAS substance database (EPA CompTox PFASMASTER), OEHHA Prop 65 weekly auto-update with diff alerts, and a tokenized B2B PFAS declaration request flow for supplier questionnaires. Just rebuilt the /learn content hub with two batches of compliance guides. Put the product through seven rounds of rigorous AI-assisted code review — security, performance, and architecture. Next up: start lining up the first paying customers.

A barbeque competition judging platform — replaces clipboards, carbon copies, and spreadsheet math with real-time digital scoring for organizers, table captains, and judges. Just wrapped a shipping-readiness pass: analytics installed, search-engine basics in place, and an automated check on every code change before it ships. Next up: a richer test setup so the app gets exercised against a real database, ahead of piloting at a sanctioned competition this season.

A marketplace for hard-to-get Taipei restaurant reservations — a scraper pulls listings from inline.app, and a modern web storefront with Stripe handles the buy side. Recent AI-assisted work has been mostly scaffolding: an admin dashboard, a way for Claude Code to poke at the scraper and database directly, and project-specific shortcuts for running tests and keeping docs in sync. Honest blocker: neither side of the transaction is live yet — checkout isn't wired and the scraper has barely been run — so the next push is getting one full listing-to-purchase flow working with real data.

Laravel 8 API backend for Tastemakers — live on Railway at api.tastemakersapp.com. GitHub auto-deploy active (push to main → Railway builds automatically). Debug/diagnostic routes removed (debug-schema, debug-signup, run-schema-fix, clear-cache). Remaining security work: unauthenticated delete endpoints and IDOR vulnerabilities. Production DB is empty — all real data still on legacy Namecheap MySQL host.

My food blog from 2007–present, rebuilt from Wayback Machine archives and Instagram exports — 1,639 posts and 1,400+ mapped restaurants live at thirstypig.com. Venue tags long-tail push ongoing: 655 venue JSONs published, 885 posts displaying tag pills, 699 venues in venues.yaml. Python test suite at 118 (+ new test_scrape_google.py for auth_gated() heuristic). Next up: keep extending venue tags (use curate_candidates.py --min-posts 1), then roll out the Bold Red Poster redesign across the rest of the site.

Wired the entire Claude-assisted project ecosystem into a single, searchable knowledge base — todos, docs, decision records, and session notes across all 8 active projects, linked and cross-referenced in one vault. Synced across Mac and iPhone via iCloud. Built a pipeline from Obsidian → GitHub → thirstypig.com for the restaurant hit list, so updating the list on my phone updates the blog automatically.

Testing an MCP server that gives Claude Code (and any other MCP-aware agent) read/write access to my Upnote notes. The pull is keeping fleeting thoughts, draft prompts, and reference notes outside the chat window but still accessible to whatever agent is doing the work — no context-switching to the app to copy a paragraph back in. Running it alongside the Obsidian setup to see whether one becomes the canonical capture tool, or whether they cleanly split (Upnote for raw thinking, Obsidian for project ops).

Using Claude Design to redesign three of my sites — The Fantastic Leagues, Aleph, and The Thirsty Pig. Of the AI design tools I was comparing (Figma AI and Lovable were the others), Claude Design earned the spot — it generates component specs and UI code I can wire straight into the existing stacks instead of scaffolding a new app.

Running OpenAI's Codex CLI agent in a second terminal tab alongside Claude Code, to see whether a multi-agent split makes sense for day-to-day work. The hypothesis: Codex is especially strong at test generation, so the experiment is "Claude writes the feature, Codex writes the tests." Early instinct for this repo's style: one agent deep beats two agents shallow. Want to run it on a harder codebase before I commit.