Latency you can feel
A cached greeting or confirmation returns before the caller finishes exhaling. No awkward gap.
VARTA Cognitive Engine
LLM-last, not LLM-first
Most voice stacks send every word to a large model. VARTA classifies, caches and retrieves first, and reasons only when nothing cheaper can answer. That is why it is fast, affordable and resilient.
- <500msEnd-to-end response
- <120msCached turn latency
- 62%Turns never touch an LLM
- 9Pipeline layers
Architecture
Three zones, nine layers
Every layer has a latency budget and a fallback. Nothing in the pipeline can block a call.
Zone 1 — Perceive
- Layer 1
Streaming ASR
Multilingual acoustic models tuned on Indian accents, with partial hypotheses emitted every 80ms.
- Layer 2
Language & code-switch detection
Detects the caller's language per utterance, not per call, so Hinglish is handled natively.
- Layer 3
Intent & entity classification
A fast classifier resolves the majority of turns without invoking any large model.
Zone 2 — Decide
- Layer 4
Response cache
Semantic cache serves known turns in under 120ms at effectively zero marginal cost.
- Layer 5
Knowledge retrieval
Company, department and flow-scoped retrieval grounds answers in your own documents.
- Layer 6
NITI policy & guardrails
Stacked company, department and flow rules decide what the agent may say and do.
- Layer 7
LLM reasoning (last resort)
Only genuinely novel turns reach a model — so you can afford a better one for them.
Zone 3 — Act
- Layer 8
Action & integration layer
Defined action first, universal API call second, graceful degradation third.
- Layer 9
Neural TTS
Natural Indic voices with cached audio reuse for repeated phrases.
Why it matters
What the architecture buys you
Cost that scales down
At a 60% cache-hit rate, blended cost per call roughly halves versus an LLM-first stack.
Better model where it counts
Because only hard turns reach the model, you can route them to a stronger, pricier one.
Graceful degradation
If a provider degrades, cached and retrieved turns keep working. The call does not die.
Policy before generation
NITI guardrails run before the model speaks, not as a post-hoc filter on its output.
Indic-first speech
Recognition and synthesis tuned on Indian accents and code-switched speech, not adapted from English.
Put VARTA on your phone lines
Talk to a live AI agent for ten minutes — no signup, no card. Or book a walkthrough with our team.