NITI: the rule engine inside VARTA
Transcript
- 0:01
Every voice agent waits for a language model. VARTA mostly does not. Meet NITI.
- 0:07
NITI is the rule engine inside layer five. Author rules, DAKSH rules and global defaults, merged in order. First match wins.
- 0:15
Each condition is a pure function of the caller's words: option matches, slot aliases, any-of keywords, regex, phonetic variants. No model, under two milliseconds.
- 0:26
The matching rule names an action: accept, reprompt, redirect, defer, confirm the value, escalate. The LLM is the last resort.
- 0:36
Global defaults cover every input type; any step can override them.
- 0:40
Every turn NITI resolves saves three to four seconds and about forty paise. So the goal is simple: maximise the share of turns that never leave NITI.
- 0:52
In the Bank KYC replay, eleven of thirteen turns never touched a language model. vartavoice dot ai.
About this film
NITI (नीति, policy) evaluates author-written and DAKSH-generated rules against the caller’s words with zero model calls: option matches, slot aliases, keywords, regex, phonetic variants. The matching rule names the action, and falling back to the LLM is the exception. In the Bank KYC replay, eleven of thirteen customer turns never touched a language model.
1:01 · · English captions
More films
Teaser · 56 s
The model can wait.
In a world where every voice agent waits for a model, one engine decided not to wait. Rules first, model last: seven layers, zero hesitation, the same voice at half the bill.
Explainer · 4 min
Train once. Run a million calls.
DAKSH uses a language model once per workflow, the way a programmer uses a compiler. VARTA runs every call on seven layers of code, asks the model last, and never lets it invent a route.
Walkthrough · 4 min
Configured once. Decided every turn.
Step by step through a bank KYC workflow: the configuration DAKSH writes at each step, and how it lets VARTA decide the call inside the engine instead of asking a language model.
Head to head · 1 min
Same call. Three agents.
One caller, one request, three voice agents: a cascaded pipeline, a speech-to-speech model and VARTA. Watch which one waits, which one mishears the address, and which one confirms it by rule.
Inside the engine · 47 s
One sentence through the VARTA engine
Follow one caller sentence through VARTA layer by layer: reflex, recognition and rules first, an LLM only when nothing else can answer, every step on one visible ledger.
For enterprise · 1 min
Why VARTA works for enterprise
Fluency no longer closes deals. Decisions by rule, writes to your ERP mid-call, live data volunteered only when a threshold is crossed, and a deterministic cost engine, in four markets.
Premium voices · 1 min
Keep your voices. Cut the bill.
Keep the premium voice you chose. VARTA replays the lines an agent has said before from cache, in the same voice, and sends only the new ones to your provider.
UPCHAAR and SHALYA · 2 min
How VARTA heals itself
Most voice agents stay as smart as the day they were written. VARTA’s two healing layers learn from every LLM rescue and propose structural fixes, checked before anything goes live.
Launch film · 24 s
Introducing VARTA
A voice-agent platform with a visible cost ledger: seven explicit layers, cached audio for repeated phrasing, and only the turns that need reasoning reach an LLM.