Broke up the GPU monolith. Lower latency, lower bill, deploys that don't need a war room.

Hyperstate AI ran an AI-assisted music production platform. Creators uploaded audio and worked with a producer-style assistant (the Louis Bell persona) that kept full context across sessions. The startup ran out of funding after launch.

A single GPU-heavy server handled audio processing, lyrics, transcription, MIDI generation, and the producer agent. Deploys were manual. The compute bill grew faster than usage. The architecture carried a demo but couldn’t carry the launch. No horizontal scaling, no failure isolation, one bad deploy took the whole product offline.

Drove the transition of GPU-heavy integrations out of the monolithic single-server setup into a distributed, microservice-oriented backend with clearly separated responsibilities (agents, audio processing, and generation services), each with its own scaling envelope. Drove the modernisation of the deployment stack from manual deploys to a dockerised, orchestrated infrastructure across every component. Drove the migration from local, self-managed, compute-heavy audio, lyrics, and transcription libraries to lightweight, scalable alternatives, cutting latency and the compute bill at the same time. On top of that: Django REST API, thirdweb-linked JWT auth, project and sample management, MIDI generation workflows, Neo4j knowledge graph, PostgreSQL. Same product surface, infrastructure that carries growth instead of fighting it.