AI projects
systems I've designed and shipped
A scheduled personal AI system that reads across Slack, Granola, Google Drive, and Obsidian to generate weekly reports -- therapy reflections, executive coaching, relationship insights, product updates, and more. Reports build on each other through tiered dependencies and shared working memory, so each one has the context of the others. Runs autonomously on macOS via launchd, with a manifest-driven architecture that makes adding new report types a config change, not a code change.
Internal tools for MILLS ("Beautiful, Buildable Bathrooms") -- a platform connecting general contractors with homeowners through design-led renovation packages. Each tool translates hard-won contractor knowledge into software: a tile engine that calculates quantities and waste by layout pattern, a spec library that generates deterministic install scopes from room dimensions, a blocking calculator that outputs cut lists from floor plans, and a real-time pricing crawler that tracks material costs across suppliers with async processing and Slack alerts.
TypeScript engine deployed from Replit. Computes tile quantities, waste factors, and layout patterns from room dimensions and tile specs.
Structured data hierarchy -- Materials / Objects / Phases / Trades / Tasks / Variables -- that enables AI-generated scopes and materials lists.
GC-validated dual-height output with the "Shim Down" principle. Converts floor plan measurements into blocking cut lists.
AI-powered materials order system that reconciles supplier confirmations against project BOMs and flags discrepancies.
Made AI tooling mandatory across a 4-person engineering team. Designed a task triage framework ('Magic Quadrant') to identify which work AI handles autonomously vs. what needs human review -- measured in story points per review hour. Set up Cursor background agents integrated with Linear for autonomous PR generation, ran team trainings on Claude and Cowork, and led a codebase-wide TypeScript strict mode migration using AI-assisted refactoring.
Built at the dawn of vibe coding (late 2024) to scratch my own itch: learning more in less time from podcasts. Grew into a platform where users subscribe to YouTube channels and receive auto-generated summaries via a mobile app or email. Shipped web and iOS -- an end-to-end automation tool from audio ingestion to formatted delivery.
The prototype that launched MILLS -- used to pitch early pre-seed investors. Explored how to translate homeowner style preferences and room dimensions into fixed-price bathroom renovation packages, validating the core product thesis before engineering began.
An early MILLS MVP (mid-2025): an N8N workflow chaining 8 LLM calls to produce a 30-page personalized renovation roadmap -- AI-generated room images, phased budget, install specs, and a 'what to expect' guide -- delivered as a PDF based on the customer's room dimensions and preferences. Built before AI-generated everything became commonplace. Replaced by more targeted tools as the product matured.