EVA — Enterprise Voice Assistant

EVA is a production-grade, multi-source enterprise reasoning system with a real-time voice interface. Customers can speak in English or Bangla, or type, and EVA retrieves live account data from six independent data sources, reasons across all evidence simultaneously, and responds with a traceable, confidence-calibrated answer rather than a hallucination.

Each turn runs through a five-stage pipeline in which the LLM is invoked exactly three times — for intent, reasoning, and call summary — while tool selection, evidence sufficiency checking and aggregation stay fully deterministic. Tools execute in parallel via asyncio.gather with independent database sessions and a three-second timeout, and a single replan pass runs when the first round of evidence is insufficient.

Voice runs on Gemini 2.5 Native Audio over a WebSocket bridge that injects the customer_id server-side at every tool call, so the model never sees the identifier. Conversation memory is PostgreSQL-backed, with the three most recent sessions injected into the voice system prompt at zero latency, and every call ends with a generated summary persisted to the database.

The system ships with an automated evaluation harness scoring 50 scenarios across five categories on retrieval, reasoning, hallucination and quality metrics, plus Langfuse tracing, Sentry monitoring and structured logging. It deploys as a five-container stack behind Nginx with Let's Encrypt SSL, built and released through GitHub Actions.