PulseFM — AI Lofi Radio
PulseFM is a 24/7 AI-generated lofi radio station: listeners vote on the genre, mood, and energy of the next track, a music-generation model renders a new song from the winning combination, and every connected listener hears it change in sync — no page refresh, no polling, no pre-rendered library.
No feature screenshots or a walkthrough video yet — these docs are text/diagram-only for now.
Live: pulsefm.app · Source: github.com/amit-00/pulseFM
Role
Solo project — designed and built end-to-end over ~215 commits (Jan–Mar 2026): the Next.js frontend, five Python backend services, four Cloud Functions, the GPU generation worker, and all Terraform infrastructure.
Tech stack
Frontend Next.js 16 · React 19 · Auth.js · Server-Sent Events Backend FastAPI on Cloud Run · Python Cloud Functions (Gen2) AI/ML ACE-Step music diffusion model on Modal · GPU memory snapshotting Data Firestore · Redis (Memorystore) Messaging Pub/Sub · Eventarc · Cloud Tasks Infra Terraform · GCP · Cloud Build · Docker
What's documented here
- Problem & Motivation — what this is solving and for whom
- Solution & Features — what it does
- Technical Architecture — how it's built, and why
- Challenges & Solutions — the hardest problems and how I solved them
- Results & Impact — what shipped, what I learned, what's next
Editorial
Rebuilding PulseFM: the clock comes first
v2 collapses five services and three functions into two, drops Redis for create-only Firestore documents, and starts from the one thing most likely to embarrass the project — a synced audio clock — before any generation code exists.
Introducing the editorial log
A place for build notes, updates, and retrospectives on PulseFM.