Call Center Compliance API

Call centre conversations have to be checked against standard operating procedures, but reviewing them by hand does not scale, and the calls themselves mix languages mid-sentence.

Who was this built for, and what was the manual process it replaced? One line on the review volume makes the case immediately.

• Mixed-language audio — calls switch between Tamil/English and Hindi/English mid-sentence, which off-the-shelf monolingual models handle badly.
• Audio over REST — clients submit base64-encoded audio to an HTTP endpoint, so payload size and request duration both bound the design.
• Custom authentication — the API is exposed to external clients and had to carry its own auth rather than sit behind someone else's gateway.

Add the real numbers: typical call length, payload size ceiling, and the latency you targeted per request.

Base64 audio is decoded and passed through speech-to-text, then an NLP and LLM layer evaluates the transcript against SOP rules, scores sentiment, and classifies mixed-language segments. Transcripts are embedded into a vector index so past calls can be searched semantically.

Name the actual components — which STT model, which LLM, which vector database — and show the request path as a diagram.

Why a vector index rather than keyword search over transcripts? Why LLM evaluation rather than a rules engine for SOP checks? Each answer with the alternative you rejected.

No measured figures exist for this project yet. Even one — SOP classification accuracy on a labelled sample, or median end-to-end latency per call — would change how this page reads.

Where does the mixed-language classifier fail? What happens on long calls or poor audio quality?