How VARTA heals itself

Transcript

1. 0:01

   Hi, I'm Anaya, a VARTA agent. Most voice agents stay exactly as smart as the day they were written. VARTA doesn't. It has two healing layers.

2. 0:11

   When my rules miss what a caller meant, the LLM rescues the turn. That costs money, and about a second and a half. UPCHAAR, which means treatment, learns from every rescue.

3. 0:24

   A small background call picks one to three phrases from what the caller actually said. It never sits on the caller's path.

4. 0:32

   Six plain-code checks run in milliseconds: grounded, specific, would have caught it, steals nothing, not a duplicate, within limits.

5. 0:42

   Pass all six, and the phrase is added to the workflow at once, for every call from its next turn. The next caller who says it is answered by rules: free, in milliseconds. Every heal has an undo.

6. 0:58

   SHALYA, which means surgery, works on structure. During the call, free checks spot gaps: a side topic no intent covers, a question asked twice, an answer that clashes with the next question.

7. 1:13

   When a flagged call ends, one LLM review turns the evidence into findings. Seen on three calls, a finding becomes a proposal, drafted as the same edit operations the Designer uses.

8. 1:26

   Before anyone accepts it, a graph check and a replay of the last two hundred calls show exactly what would route differently. A person accepts, and every call from the next one runs the new version.

9. 1:39

   Nothing structural changes without a person. Automatic mode accepts only matured, low-risk proposals, then watches the next twenty calls and rolls back on its own if completion drops.

10. 1:50

    Every call leaves the workflow cheaper and more correct, with a history you can read and reverse. That is healing, the VARTA way.