Roboscribe
Plant inspection and maintenance walkthroughs, captured and made searchable. Multilingual site recordings - diarised, transcribed, translated Indic-to-English, and enriched against open maintenance tickets - running fully offline on an air-gapped host.
- Role in stack
- Plant capture -> searchable evidence
- Current state
- Working code
- Best demo use
- Plant/facility maintenance capture for anyone evaluating the maintenance growth area
- Honest boundary
- Calibration figures are not yet settled; accuracy depends on audio quality and language mix
Eight stages: audio extraction, NeMo diarization, short-segment merge, language-routed ASR, IndicTrans2 translation, Qdrant RAG enrichment, a reviewer sign-off gate, and the run manifest. Completing the run opens a transcript explorer with speaker and language filters, text search, and per-segment translation and glossary detail.
The short-segment merge stage is the real story: diarization over-segments, and Whisper hallucinates on fragments under ~3s. The demo reproduces that failure and shows the fix.
The demo is a static mock. It runs entirely in your browser on synthetic data — nothing is gated, nothing is collected.