Things I've built

Problem

Manufacturers were reading dense engineering drawings by hand to pull out parts and build quotes. It was slow, inconsistent, and impossible to scale.

Approach

I designed the whole pipeline. An OCR layer (Docling and PaddleOCR) lifts text off large drawings, a real time YOLO model detects and classifies components, and LLMs turn the mess into structured product data that feeds an automated quoting flow. I also built the eval harness that kept model quality honest.

Problem

Matching or rebuilding a site's look is guesswork. People eyeball colors, fonts and spacing and still get it wrong.

Approach

DesignMD crawls a site, single or multi page, and pulls the real design tokens: color, type, spacing, shadows, components. It emits a structured DESIGN.md spec plus an HTML preview you can build against. I built the extraction engine, the CLI, and an eval harness that scores how faithfully it captures a brand.

Problem

Connecting musicians means matching and searching across a large, always growing graph of people and work, without the app slowing down.

Approach

I led the backend architecture: scalable microservices with ElasticSearch driving matching and search on top of PostgreSQL as the source of truth. Built to keep discovery quick as the user base grew past six figures.

Problem

Good prompts get scattered across ChatGPT, Claude, Gemini, Perplexity and Grok, with no easy way to keep or reuse them.

Approach

A Manifest V3 Chrome extension that captures prompts from five AI platforms, then organizes them locally with folders, tags, fuzzy search and Markdown or JSON export. Everything runs in the browser. No account, no backend, nothing leaves your machine.