Stack and request flow - RAWCHUNKS

The browser handles interactive editing. Fastify handles requests and project records. Python workers run audio models on NVIDIA L4 GPUs. Hosted models prepare briefs, lyrics, patterns and artwork.

Components

Technology	Role in Rawchunks
React, TypeScript and Vite	Studio forms, Lyrics Canvas, library and Editor UI.
Zustand	Timeline, selection, track, mixer, instrument and Code state.
Tone.js and Web Audio	Schedule tracks and synthesized parts; process effects; render merged audio locally.
Strudel notation	Express rhythmic and pitched patterns. Rawchunks parses its supported notation and schedules it in the browser.
Node.js, Fastify and Zod	Serve REST and NDJSON, validate inputs, admit paid operations and coordinate results.
Python and PyTorch	Run ACE-Step conditioning, diffusion, waveform decoding, separation and voice training.
vLLM	Serve ACE-Step’s music language model alongside the audio pipeline.
TorchInductor and Triton	Compile PyTorch operations into GPU kernels. Compiler caches and warmups prepare the inference paths.
NVIDIA L4	GPU compute for each audio-model or training replica.
Kubernetes on GKE and KEDA	Run separate model pools and adjust replicas from pending queue length.
Valkey	Durable jobs, executor leases, progress messages, WIP state and shared caches.
Neon PostgreSQL and Drizzle	Store accounts, songs, takes, stems, albums, history and Credits records. The API uses Neon’s HTTP driver for project queries.
Cloudflare R2	Store audio, artwork, voice samples and generated media.
Nebius Token Factory	Serve the Nemotron models used by both agents and image-brief interpretation.
Cartesia and ElevenLabs	Transcribe brief recordings and recover or align song lyrics.
Google Nano Banana	Generate cover artwork.
USDC sessions, x402 and Credits	Fund music operations and listening. Payment interfaces.
Thirdweb and Swig	Connect EVM and Solana accounts for the app’s wallet workflows. Account setup.
IPFS and chain contracts/programs	Store mint metadata and record music ownership, licenses and royalties. Publishing.

Connections

Audio submissions return job IDs so the API can keep serving other requests. Progress reaches clients through NDJSON or status reads. Completed model output becomes a saved master, candidate take or set of stems. Concurrency covers scheduling. Inference covers the work inside each GPU replica.