Projects

AI testing companion trained on 928 real JIRA tickets from healthcare products (ticket metadata only, no PHI). Hybrid model: LSTM neural network plus vector similarity search. Predicts bug severity, suggests test cases, flags regression risks. Gamification (XP/levels/achievements) and 5 AI coaches make testing less boring. Turns out personas that roast you for skipping edge cases actually work.

Clicking through JIRA tabs to review 20+ tickets per sprint was killing me. Built a wrapper with 5 views (board, list, tree, detail, calendar) and full keyboard navigation. No more tab-clicking hell. React Query caching, loads 50 tickets in under 2 seconds.

Productivity platform for ADHD and executive dysfunction. I've tried (and abandoned) Todoist, Notion, Things, Obsidian. None of them work for ADHD brains. So I built my own.

12 features: task dependencies, habit streaks, mood tracking (spoon theory), journaling, Google Calendar sync, concert tracking, RSS reader, multi-pane workspace (Andy Matuschak-style), dashboard, analytics. ADHD-specific UX: progressive disclosure, color-coded priorities, celebration animations.

Forever project, rebuilt multiple times. Each iteration shares the same Supabase backend (row-level security), so I see my real data while building. Building IS the accountability system.

Timeline-based memory system for tracking events, notes, and context over time. Key-value storage for quick lookups, MCP server integration for Claude, full-text search, tag filtering. Built with TanStack Table for timeline visualization and Express for the API backend.

Rebuild times hit 12 minutes and I couldn't iterate fast enough. Honestly unbearable when you're trying to test a one-line CSS change. Built a monorepo with Turborepo managing 16 apps. Shared UI components, cached builds.

Turborepo handles parallel builds and incremental rebuilds. First build ~15s, cached rebuild under 200ms. Down from 12 minutes when I started.

Coordinates 5 AI agents with isolated workspaces. Gateway routes messages to specialists: one handles tasks, one does character narration, one manages scheduling. Each agent only sees what it needs.

Tiered model selection keeps costs low: simple stuff uses DeepSeek Free, longer conversations use V3.2, complex reasoning uses Claude Haiku. Monthly cost: ~$5-15 instead of $100+. Local HTTP gateway handles token auth, session management, and inter-agent relay for time-sensitive tasks.