Miles Donovan Bernhard

An agent reads your documents and feeds into a book of transactions, proposing accounts, tags and groupings, and drafting an extractor when a layout is new. Everything it learns lands in your repo as plain text: the rule that tags a merchant, the SQL behind a chart you asked for once and kept. It proposes; you settle. Every account is checked against the balances its statements print. The book is CSV in git; the SQLite read model is derived and rebuildable. Python on FastAPI, a React PWA with a chat drawer, and an MCP server, all on one action API — the agent can do anything the screens can. Runs on your machine or your own cloud.

A globe of every place a Chase Sapphire Reserve credit reaches: the ~1,360 hotels in The Edit, flagging the ones where the Select Hotels credit stacks for $500 off a stay, plus the ~270 restaurants in the $300 Exclusive Tables dining credit. Hotel coordinates are recovered from a community map's viewer payload; restaurants are scraped from OpenTable with a headless browser, each carrying exact coordinates, price band, and rating pulled from its profile. A weekly job rebuilds both lists and reports what changed. Live hotel rates run on each visitor's own API key — encrypted in their browser, proxied once, never stored — on Cloudflare Workers and D1.

A Model Context Protocol server for YNAB that lets AI assistants read and manage your budget through natural language — 13 focused tools, a stateless "Sign in with YNAB" OAuth broker, read-only mode, and one-command Docker deploys for self-hosting.

A CLI, published on npm, that reconciles YNAB account balances against SimpleFIN Bridge — it fetches what each account is actually worth and posts a single balance adjustment per account. Guided setup, dry-run previews, and CI-tested. Or skip the CLI: the hosted app installs a GitHub App onto one repository you pick — it asks for no administration scope, so you create the repo and it touches nothing else — then provisions the scheduled sync there, with credentials sealed in your browser and written straight to that repo's encrypted secrets.

An app for creating, discovering, and reviewing music events — from dorm parties to festivals. Deep market analysis, wireframes, a product manual, and an MVP, built through HMC INQ, Harvey Mudd's alumni-mentored incubator.

A web app that colorizes black-and-white images with a convolutional neural network, built for CSCI121: Software Development at Harvey Mudd using Flask and OpenCV.

A year-long research project through the Gould Center for Humanistic Studies tracing hip-hop's trajectory geographically — data collected via the Spotify and Genius APIs plus historical Billboard Hot 100 charts, paired with qualitative analysis.

A seeded generative art piece built in p5.js for CSCI198 at Harvey Mudd — enter any seed and it renders a deterministic 3D composition mapped onto a globe, so the same seed always produces the same artwork.