From sound to something useful.

1. Upload without blocking

The Next.js frontend asks FastAPI to create an upload. FastAPI creates a private object key and a multipart upload in AWS S3. The browser uploads up to three 16 MiB parts concurrently using short-lived signed URLs. Each part retries up to three times if the connection fails.

The API checks the part count and total byte size against storage before completing the upload. Large audio never passes through a Next.js request body. The default configuration permits files up to 50 GiB, with no duration cap; available disk space, processing time, and provider credits remain practical limits.

2. A durable background queue

Completing an upload inserts a job and updates the recording in one PostgreSQL transaction. The HTTP request returns immediately. A separate Python worker claims a job using FOR UPDATE SKIP LOCKED, so multiple workers can safely consume the queue.

A worker renews its lease every 15 seconds. If it crashes, another worker can reclaim the job after 180 seconds. Every write checks lease ownership. Transcript sections are checkpointed, so a retry skips sections already saved.

3. Handling long audio

FFprobe checks that the file contains decodable audio. FFmpeg converts one section at a time to mono, 16 kHz PCM WAV. Each section is at most 30 seconds long, with one second of overlap to reduce words lost at a boundary.

Sections go to Gnani’s REST API sequentially, below its 60-second request limit. Exact repeated word sequences at boundaries are removed before the transcript is assembled. Timestamps refer to chunk boundaries, not word-level alignment. The worker downloads the original to temporary disk and removes temporary files when it finishes or fails.

4. Grounded summaries

Gemini receives the transcript, not the audio or the API key. Its instructions request an overview, key takeaways, topics, and only explicitly stated action items. Responses use a JSON schema and are validated before saving.

Long transcripts are split into bounded text sections, summarized individually, and recursively combined. The transcript is saved before summarization starts. If Gemini fails, the user can read the transcript and retry the summary without paying to transcribe again.

What lives where

PostgreSQL: recording metadata, processing states, transcripts, summary JSON, section checkpoints, and durable job leases.

Object storage: the original audio in a private AWS S3 bucket in both local development and production. Signed URLs grant temporary upload and playback access.

Browser: a signed, HttpOnly workspace cookie and the preferred library view. Audio and transcripts are not stored in localStorage. Separate browser profiles have separate libraries; clearing the cookie loses access to that workspace.

Visible progress and failures

Upload progress comes from actual bytes transferred. During processing, the frontend polls saved state every 2.5 seconds. Transcription progress advances when a section is committed; queued and summarization states explain what is happening without pretending to predict an exact completion time.

Provider timeouts, rate limits, invalid audio, missing configuration, and storage outages surface as actionable messages. Provider calls use bounded retries with backoff. A failed recording retains its original audio and completed transcript sections for an explicit retry.