VaaniDesk

VaaniDesk is a production-oriented multilingual AI customer support platform. It combines model intelligence with controlled tool execution, hybrid retrieval with source citations, evaluation coverage, and security controls suitable for real operational workflows.

Customer support systems that bolt a chatbot onto an API often fail in production: uncontrolled tool calls, weak language coverage, missing citations, no evaluation harness, and thin security around sensitive actions. VaaniDesk addresses those failure modes as first-class engineering constraints rather than afterthoughts.

The system is organized as layered services: client interfaces, API and agent orchestration, an intelligence layer for multilingual routing and retrieval, and a data/infrastructure foundation on PostgreSQL, Redis, and Docker.

Language support covers English, Hindi, Hinglish, and Marathi, with routing and evaluation awareness for mixed-language customer interactions common in real support environments.

Agent behavior is constrained through structured workflows and controlled tool calling. Sensitive or destructive actions require confirmation rather than unconstrained autonomous execution.

The platform integrates business tools behind a gated execution model so the agent can act on real operational tasks while remaining within authorization and confirmation boundaries.

Retrieval combines hybrid search over PostgreSQL with pgvector, returning source citations so answers remain inspectable. This keeps model outputs grounded in indexed business knowledge rather than unconstrained generation alone.

Security controls include authorization checks, idempotency for side-effecting operations, and prompt-injection defenses. Security-critical evaluation cases are part of the verified release metrics.

Optional STT/TTS paths exist for speech-related support flows. Real external STT/TTS providers are credential-dependent; deterministic providers/simulators are used for local and CI verification. This is not a dedicated vision or image-analysis pipeline.

Evaluation is treated as a release gate. Verified v1.0.1 metrics include 113 deterministic evaluation cases passed, with 40 security-critical evaluations and 0 security failures.

Operational visibility is built into the architecture so agent runs, tool calls, and retrieval behavior can be inspected during development and verification.

The verification suite spans backend unit/integration tests, deterministic evaluation cases, Playwright end-to-end coverage, static typing with mypy, Docker health checks, migration cycles, seed idempotency, and secret scanning.

Backend tests

206 passed · 0 failed · 0 skipped

Deterministic evaluations

113 passed

Security-critical evals

40 · 0 failures

Playwright E2E

14 passed

mypy

Clean across 100 source files

Release checks

Docker · migrations · seed idempotency · secret scan

Python · FastAPI · PostgreSQL · pgvector · Redis · Next.js · TypeScript · Docker · pytest · Playwright

• Real external LLM, STT/TTS, SMTP and WhatsApp integrations are optional and credential-dependent; deterministic providers/simulators are used for local and CI verification.
• Credential-dependent integrations are not presented as live in every deployment environment.
• An MCP server and a dedicated vision / image-analysis pipeline are not part of the shipped v1.0.1 surface.

Shipping an AI support system as engineering work means treating evaluation, authorization, idempotency, and language edge cases as part of the product surface — not as demos bolted onto a model call. Controlled tool execution and hybrid retrieval with citations were essential to making agent behavior inspectable and safe enough for operational use cases.