Enterprise Voice AI Platform · Technical Preview

LLMs understand what the customer says.
Vaartalaap controls what happens next.

A deterministic execution layer for enterprise voice AI. Graph-based routing, contract-enforced state, and hybrid TTS caching make voice automation predictable, testable, and economical at scale — without giving up what LLMs are actually good at.

IN OK Customer Deterministic graph
~75%TTS latency reduction via hybrid caching
15+intents handled in reference bank-collections bot
10static validation checks, zero I/O, before any graph deploys
Solofounder-built, self-funded, no external capital
The Problem

Enterprise voicebots fail in production in specific, recurring ways

Not hypothetical — observed firsthand building and operating voicebots for regulated enterprise customers.

High latency

Every response generated by an LLM at runtime — no caching, no pre-computation.

Non-determinism

The same utterance can produce different responses on different calls — unacceptable in regulated industries.

Development doesn't scale

A prompt change for one flow can silently alter behavior in another — a butterfly effect, not modular.

Testing doesn't scale

No way to turn "does the model do the right thing" into a systematic regression suite.

Pronunciation is unsolved

Generic TTS mispronounces exactly what an enterprise cares about — its own brand and product names.

Vendor lock-in & cost

Runtime generation via a single LLM provider makes cost grow linearly with volume, with no way to optimize.

The Architecture

Deterministic-first, not LLM-first

The LLM is invoked once per turn to classify intent. A statically-validated graph — not the model — decides what happens next.

Speech STT pluggable LLM classify intent, once Graph Engine deterministic routing contracts + validation Templates no LLM involved TTS out

Traditional LLM-first

  • LLM decides routing → non-deterministic, hard to audit
  • LLM generates responses → high latency, high cost
  • Full TTS synthesis every time → 400–800ms per utterance
  • One prompt change → butterfly effect across all flows
  • Single LLM provider → vendor lock-in

Vaartalaap

  • Graph engine decides routing → auditable, compliance-controlled
  • Templates render responses → consistent, testable, no LLM involved
  • Hybrid TTS cache → ~75% latency reduction
  • Flows are YAML → no butterfly effect from prompt edits
  • Every AI layer pluggable → no vendor lock-in
Differentiators

Five things that are implemented, not theoretical

3.1

Deterministic routing, LLM-only classification

The LLM answers one question per turn — "what did the customer mean?" — and returns a named intent. The graph, not the model, decides what happens next. Given a sequence of intents, the output is completely predictable.

3.2

Contract-enforced state with static validation

Every node declares what it requires, what it produces, and its success condition. A pure-Python static validator checks every graph before it can be saved — a structurally defective graph simply cannot exist.

3.3

Hybrid TTS — pre-generated static + dynamic stitching

Most of what a bot says is identical across calls. Static segments are pre-synthesized at startup; only customer-specific tokens are synthesized live. Measured: ~600ms → ~150ms per utterance.

3.4

Vendor independence as an architecture constraint

LLM, STT, TTS, and telephony each sit behind a configurable interface. A live Model Settings console lets an operator switch providers for a running deployment with zero code changes and zero restart.

3.5

YAML graph DSL — conversation logic as auditable data

Conversation flows are data, not code. Compliance teams can review the graph itself — not a black-box prompt — for certification.

Current Validation

Exercised as a working system, not just designed

CapabilityStatus
Audio in → audio out (full pipeline)✔ Working end to end
Bank collections reference bot (15+ intents)✔ All paths exercised
Hybrid TTS cache✔ ~75% latency reduction, measured
Dual-LLM failover (Mistral → Nemotron)✔ Automatic on primary failure
Twilio WebSocket telephony transport✔ End-to-end phone call completed
Barge-in detection✔ Silero VAD stops TTS mid-playback
Automatic call quality scoring (Auto QA)✔ Runs on every completed call
Live provider hot-swap (STT/LLM/TTS)✔ From Studio, no restart
PostgreSQL live call store✔ Calls, turns, slots, QA, analytics
Target Industries
Banking & Financial Services Insurance Healthcare Telecom Government
About the Founder

Shubham Singh

Founder & Engineer, Vaartalaap

Built while working on an enterprise voicebot in a regulated industry — an entirely LLM-driven stack that worked in demos and failed in production in ways that were extremely difficult to debug or audit. Vaartalaap is the architectural answer: a deterministic graph engine, contract-enforced state, and static validation, designed and built solo, self-funded, with no external capital raised to date.

shubham.subodh@gmail.com · Source code available under NDA on request