Pitch Coach AI
Real-time AI sales coaching with two-brain architecture for sub-second voice latency and async rubric grading.
2026
System Architecture
The Problem
Sales reps need realistic practice against varied buyer personas, but AI voice latency >1s breaks immersion and real-time rubric grading takes 5-15s — impossible within the latency budget.
Approach
Two-Brain Architecture — Brain 1 (gpt-realtime-1.5 + LiveKit WebRTC) handles real-time voice at <1000ms, Brain 2 (o4-mini + Azure Functions) handles async rubric grading, connected via Azure Blob Storage "air gap."
Architecture
Brain 1 uses audio-to-audio modality (no STT/TTS transcoding tax) via LiveKit Cloud's managed SFU. After session ends, transcript JSON triggers Brain 2 via Blob event, which grades against ISR (90-pt) or OSR (100-pt) rubrics using Structured Outputs, then emails coaching reports via Azure Communication Services.
Results
- 12 Architecture Decision Records
- 199+ pytest tests with rubric calibration validation
- Dual-rubric scoring system (ISR 90-pt / OSR 100-pt)
- Snowpipe → Snowflake analytics pipeline
- Sub-second voice latency achieved
Lessons Learned
Decoupling speed-critical and intelligence-critical paths into separate systems with event-driven coordination is the only viable pattern when latency budgets conflict with processing requirements.