Case file 01 · Conversation Intelligence
Client anonymizedRuns on site
Every call scored on one machine, inside the building.
- 01 · Context
- [Industry and size, anonymized]. Supervisors listened to a small sample of calls. The rest were never reviewed.
- 02 · Constraint
- Recordings could not leave the building. The data terms included a processing agreement, no-access clauses and deletion certificates.
- 03 · Build
- Local transcription with faster-whisper and speaker separation, then a quantized local model that scores every call against the client's scorecard. Models load one after another, so the whole pipeline fits on a single GPU.
- 04 · Result
- [N]calls scored a month[N]%of calls reviewed, up from a sample[N] hof manual QA saved a week
Published once the client approves the figures.
- 05 · Stack
- faster-whisperSpeaker separationQuantized local language modelSequential GPU loadingScorecard templates
- 06 · Timeline
- [N] weeks from scoping call to acceptance.
- 07 · Price band
- [Price band]. The list band for this package is 24,900 to 49,000 EUR.
- Hardware
- [Workstation and GPU]
- Location
- On the client's premises
- Data leaving site
- None