Bongbetic
VnAAI
An AI transcription and feedback workflow that helps trainers give sharper coaching guidance to voice agents.
AI
Training
Voice Operations
Problem
Trainer feedback cycles were slow and inconsistent across large sets of call interactions.
Outcome
VnAAI gave trainers structured transcripts and insight cues to make feedback faster and more objective.
Technical Spec Sheet
VnAAI engineering blueprint
Platform
Tauri desktop coaching assistant for call assessment
Architecture
- Frontend (Vite + React + Zustand) manages transcript upload, rubric scoring, feedback review, and settings.
- Provider abstraction (`src/lib/ai-provider.ts`) routes analysis + transcription independently across Gemini/OpenAI/Ollama/Custom/ElevenLabs.
- Persistent local store (`zustand/persist`) keeps assessments, scripts, and provider settings.
- Tauri shell (`src-tauri`) packages desktop runtime and native filesystem/dialog capabilities.
Stack
- Tauri 2 + Rust host
- React 19 + TypeScript + Tailwind + Radix UI
- Zustand persisted state
- @google/generative-ai + OpenAI-compatible APIs + ElevenLabs Scribe
- docx + mammoth for report/script document operations
Core Modules
- `ai-provider.ts`: multi-provider chat/transcription, JSON cleaning, score sanitization, fallback narrative generation
- `store.ts`: settings lifecycle, assessments history, script management, processing states
- `rubric.ts` + script parser modules: rubric normalization and custom-script-driven evaluation context
- `export-docx.ts`: assessment report generation for trainer handoff
Data Flow
- Audio upload or transcript input enters analysis flow.
- Transcription provider resolves independently from analysis provider (same-as-analysis or dedicated override).
- AI output is normalized into strict `AssessmentResult` schema with score clamps and consistency guards.
- Result surfaces in scorecards + coaching suggestions and is persisted for later retrieval/export.
Integrations
- Gemini models (analysis/transcription)
- OpenAI chat + Whisper transcription
- Ollama local model runtime (analysis)
- ElevenLabs Scribe for diarized transcription + derived voice metrics
Packaging and Delivery
- Tauri desktop bundle with Windows installer artifacts in `src-tauri/target`
- Vite build pipeline for frontend static assets
- Runtime settings-driven provider switches without rebuild
Quality and Reliability Measures
- Score validation pipeline enforces max bounds + total-score consistency + script-adherence sanity checks
- Fallback generation for missing AI fields prevents partial result renders
- Provider health testing and model-list discovery APIs in settings workflow