Company2 min read

Internal Productivity & Biometric Suite

Client-side media processing plus a face verification system for Kiosk terminals.

Next.js
Capacitor
FastAPI
Qdrant
FFmpeg
IndexedDB

An internal suite in two parts: in-browser media processing, and a face verification system running on Kiosk terminals.

Client-side media processing#

The requirement: users upload short videos that must be compressed before sending. Compressing server-side would mean pushing the full-size original over the weak connections found at points of sale.

The answer was ffmpeg.wasm running in the browser, inside a Web Worker so the UI stays responsive:

workers/transcode.worker.ts

self.onmessage = async ({ data }: MessageEvent<TranscodeRequest>) => {
    const ffmpeg = await loadFFmpeg();
 
    await ffmpeg.writeFile('in.mp4', data.bytes);
    await ffmpeg.exec([
        '-i', 'in.mp4',
        /* CRF 28 + a fast preset: small enough to upload, clear enough to review. */
        '-vcodec', 'libx264', '-crf', '28', '-preset', 'veryfast',
        '-vf', 'scale=-2:720',
        'out.mp4',
    ]);
 
    const output = await ffmpeg.readFile('out.mp4');
    /* Transferable: hand over the buffer instead of copying it to the main thread. */
    self.postMessage({ bytes: output }, [(output as Uint8Array).buffer]);
};

Face verification#

The biometric half matches faces via embedding vectors stored in Qdrant. Source images are never retained:

api/verify.py

@router.post("/verify")
async def verify(payload: VerifyRequest) -> VerifyResponse:
    embedding = encoder.encode(payload.frame)
 
    hits = qdrant.search(
        collection_name="faces",
        query_vector=embedding,
        limit=1,
    )
 
    # Threshold set high: a false reject is an inconvenience, a false accept is not.
    if not hits or hits[0].score < MATCH_THRESHOLD:
        return VerifyResponse(matched=False)
 
    return VerifyResponse(matched=True, subject_id=hits[0].payload["subject_id"])

Outcome#

Video compressed on the user's own device, cutting upload size dramatically
UI stays responsive throughout, thanks to the Worker
The verification system never stores raw face images