Prompt Pilots UTHM hits 1k agent calls/week in prod
the voice-agent stack that `@hafiz.fskm` and the circle have been iterating on for 4 months is now in production for a real UTHM Library use case.
the Prompt Pilots UTHM circle has been quietly iterating on a voice-agent stack for a library-helpdesk use case. this week, the stack crossed 1,000 real patron calls in a single week, with a 92% first-call-resolution rate.
the customer is the UTHM Library Digital desk, handling 60k patron interactions a year. before the agent, their support queue had a 14-hour average wait time. after: 47 seconds.
what the stack looks like
- whisper.cpp for STT, running on a 3090 in the FSKTM GPU cluster.
- custom Bahasa Melayu + Manglish phoneme dictionary (the off-the-shelf Whisper gets Bahasa 78% right; ours gets 94%).
- agent loop in Python with a strict tool-use schema. no free-form function calling, no MCP.
- escalation to a human librarian on a 4-of-5 confidence threshold.
what is next
we are publishing the Bahasa Melayu phoneme dictionary as a separate OSS project next month. it is the most-requested piece of the stack. if you are working on Bahasa STT and want to compare notes, hit us up.