Saepul MalikAI Call-Center Agent Copilot
Real-time transcription and a live checklist that guide call-center agents mid-call, then auto-score every session against a rubric.
Problem
Prior QA only sampled a fraction of agent–customer calls for manual review — roughly 10 out of every 100 sessions across all ~11 agents — leaving most calls unchecked. The product team wanted agents guided in the moment, not just graded after the fact.
Architecture
Live calls are transcribed with OpenAI's real-time transcription model. Per-case knowledge — what should be asked or checked for a given call type — is stored in a data table and pulled in by the AI to drive both the live checklist and the post-call evaluation, rather than relying on one generic prompt for every call type.
Execution
The idea came from 2 people on the product team; I built the full feature set end to end — live-call probing (prompting the agent on what to ask next), live analysis of the ongoing call, and QA evaluation that scores each session after it ends. Two hard problems stood out: transcription accuracy, particularly separating the agent's voice from the customer's amid background noise, and getting the AI's rubric-based scoring to land close enough to a human QA reviewer — some subjective rubric dimensions still carry a scoring margin of error above 5.
Impact
Moved QA coverage from a ~10% manual sample to 100% of sessions scored automatically, across all ~11 agents. Still testing-scale as a proof of concept, not yet rolled out product-wide.