
The problem. In India, mis-selling of insurance and credit is routine. An advisor says "the returns are guaranteed, twelve percent", the customer signs, and the only record is the advisor's own tick-box form. IRDAI logged 26,667 mis-selling grievances against life insurers in FY25, up 14 percent. The insight. The moment that matters is the sentence itself. Hear it, know who said it, answer out loud before the customer believes it, and the sale is corrected instead of investigated. Leave a record a stranger can verify in a browser, and the customer need not be believed; they can be checked. How it works. Saakshi sits between advisor and customer. AssemblyAI Streaming STT (Universal-3.5 Pro, speaker labels, English and Hindi in one sentence) is its ears. Every turn is checked against a protocol pack of required disclosures and prohibited claims, each with its regulator citation; two packs ship (IRDAI ULIP sale, RBI KFS loan). A confirmed prohibited claim becomes a scripted correction under twenty words on the AssemblyAI Voice Agent API, spoken 1.4 s (median) after the advisor stops. Then she asks three to five teach-back questions from what was said, hears the Hinglish answers and records a verdict each. Every turn is hash-chained; the public verify page reads VALID, or TAMPERED after one changed character. Evidence you can trust. Saakshi speaks to the customer, not into the seller's ear, and speaks rarely: one intervention in 36 advisor turns. The recogniser is never biased toward the phrases the rules match, so a mishearing cannot invent a disclosure (0 false ticks of 12 live). The LLM (Groq) is advisory and can never tick, flag or speak. Quotes carry no phone, Aadhaar, account or PAN numbers. Business. 5.8 million licensed selling seats in India, all owing the same disclosures. Per-seat software on the laptop already on the desk; the rulebook is a JSON file. One person can run the demo. No customers yet; every number on /metrics carries its command.
30 Sep 2026