Built on Sarvam AI · India-first speech & language stack

Voice & language infrastructure,
built Sarvam-native.

Two production-grade projects: an open-source MCP server that hands any agent Sarvam's full API surface, and a real-time code-mixed voice agent for banking & financial services.

vaani · live call req · collections-line / +91••••4412
STTSaaras v3 · "Mera EMI kitna bacha hai?"
Routeintent=account_action → Sarvam-105B
RAGpgvector top-k · clause 7.2 · grounded ✓
Toolget_emi_details() → ₹4,250
Guardwaiver request · confidence 0.41 < 0.6
TTSBulbul v3 · anushka · streaming
HITL escalate_to_human() warm transfer
Saaras v3 STTBulbul v3 TTSSarvam-105BSarvam-30BMayura translateParse · docsMCP toolsRAG · pgvectorEvals-firstPII redaction Saaras v3 STTBulbul v3 TTSSarvam-105BSarvam-30BMayura translateParse · docsMCP toolsRAG · pgvectorEvals-firstPII redaction
0
typed MCP tools
0
milestones shipped
0
test coverage
0
latency budget
Two projects

An MCP server and a voice agent

One gives every agent Sarvam's tools. The other puts them to work on a live, low-latency phone call.

01 · Open-source infra

Setu

Sarvam MCP Server
FastMCPPython 3.11+Pydantic v2OpenTelemetry
  • Exposes Sarvam speech, translation, chat & document tools to any MCP client — Claude Desktop, Cursor, LangGraph.
  • Typed Pydantic in/out on every tool, with retries, rate-limiting and a mock mode.
  • Per-call cost (INR) and latency emitted as OpenTelemetry span attributes — every call is observable and priced.
02 · Flagship demo

Vaani

Multilingual Voice Agent for BFSI
PipecatSaaras v3Sarvam-105BBulbul v3
  • A caller speaks Hindi / Hinglish; the agent answers grounded questions over policy docs in real time.
  • Calls backend tools (account status, EMI, callbacks) and escalates to a human on low confidence.
  • Evals-first: WER, tool-call accuracy, groundedness, p50/p95 latency & cost per interaction.
One turn, end to end

From a caller's words to a grounded, priced answer

Every turn runs through the same disciplined pipeline — instrumented, validated, and escalated when unsure.

01
Listen

Saaras v3 transcribes code-mixed audio; the router picks an intent and a model tier.

02
Ground

RAG retrieves policy context; a groundedness guard blocks any answer not supported by it.

03
Act

Sarvam-105B calls validated backend tools; sensitive actions route to a human.

04
Speak

Bulbul v3 streams the reply; cost, latency & PII-redacted memory are written.

typed everywhereobservabletestedreliableevals-firstcompliantcontainerisedCI green typed everywhereobservabletestedreliableevals-firstcompliantcontainerisedCI green
The Sarvam stack

One provider, the whole pipeline

Speech-to-text, reasoning, text-to-speech, translation and document parsing — all Sarvam models.

Saaras v3
STT · codemix
Bulbul v3
TTS · streaming
Sarvam-105B
agents · tools
Sarvam-30B
fast turns
Mayura v1
translation
Parse
doc → JSON
sarvam-m
24B · open
Lang ID
detect script
Engineering posture

Built like production, not a prototype

Typed everywhere
pydantic v2 · mypy strict
Observable
OTel traces + cost/latency
Tested
pytest · 80%+ core
Reliable
retries · rate-limit · mock
Evals-first
WER · groundedness
Compliant
PII redaction · audit trail
Containerised
Docker · compose
CI green
GitHub Actions