Building Mozart AI
Founding product engineer on a browser-based AI music studio. The work spans the whole stack; this is the shape of it.
Product engineering at speed
End-to-end ownership of user-facing surfaces on a product serving 250K+ people: roughly 140 pull requests, 650+ production commits, and 1,600 files across a TypeScript monorepo in four months, including a full redesign of the flagship library shipped to production. Design, build, instrument, ship, measure, iterate.
Frontend at scale
Large React and TypeScript applications with strict performance discipline: state architecture across dozens of Zustand stores, render-budget profiling, canvas and WebGL rendering for waveforms and timelines, and a design-system-first approach where every surface composes shared primitives.
Real-time and distributed systems
Multiplayer collaboration on CRDT infrastructure: presence, conflict-free sync of tracks, mixer state, and MIDI, server-authenticated sessions, durable persistence across Postgres and Mongo, and streaming audio delivery. Systems designed to fail closed and degrade to zero cost when disabled.
AI and ML engineering
Generation pipelines over multiple music models, a comparative evaluation harness with objective metrics (CLAP, SI-SDR, FAD) and human preference ranking, real-time voice agents combining streaming STT, LLMs, and TTS in acoustically hostile environments, and persona systems from training flow to inference.
Audio engineering
Studio-grade audio in the browser: a WASM DSP engine, a native C++ bridge hosting professional VST3/AU plugins, offline render and export pipelines, waveform analysis and rendering, and latency-sensitive playback architecture shared across web and mobile.
Backend and infrastructure
Cloud functions and typed service APIs, webhook-driven persistence, access-controlled CDN signing, storage architecture across GCS and CDN layers, and CI/CD across a many-package monorepo with preview environments per branch.
Data and experimentation
Typed event contracts as engineering artifacts: activation funnels, outcome and latency measurement on every user action, feature-flagged rollouts, A/B experiments on onboarding and growth surfaces, and the unglamorous work of making existing numbers honest before trusting them.