AI Transcription Platform

Problem

Anything that puts an AI provider behind a public URL is a spending endpoint. The failure is never the model, it is the accounting: a rate limit that resets when the visitor clears localStorage, a budget counted after the call instead of before it, a session that dies mid-request and leaks its reservation forever, a provider key handed to the browser because that was easier. This is the same problem a client has in production, with my own invoice as the test.

Approach

• →Spend is reserved before the provider is called and reconciled against the real duration when the session ends, so a request can never start against a budget that is already gone. Reservations that are never finalized expire on a sweep instead of being lost.
• →A daily ceiling per cost category on UTC day boundaries, plus a kill-switch environment variable that refuses new work immediately, with no redeploy.
• →The browser never sees the account key. Live transcription mints an ephemeral Deepgram key with a 5-second TTL and the narrowest scope that still opens the socket.
• →Rate limiting keys on the real TCP peer that Traefik sets, not on a client-controllable header. Callers arriving without it are pooled into a single hostile bucket rather than each getting a fresh allowance.
• →Provider errors are sanitized before they reach the logs, so a failed call never writes a key or a token to disk.
• →On top of that, three input pipelines: file upload sniffed by magic bytes, YouTube via yt-dlp with IPv4-pinned clients to survive datacenter-IP blocking, and live microphone. An optional LLM pass reuses the cached transcript instead of paying to transcribe twice.

Stack

By the numbers