TRACE at GyanDhan — 2,000+ WhatsApp groups, fifteen minutes a day.

GyanDhan runs an education lending platform with partners across India. Most of the partner communication happens on WhatsApp. Before TRACE, the ops team spent 2.5 hours a day handling queries across 2,000+ active groups. After TRACE, it's 15 minutes.

The problem

Each partner is in a WhatsApp group. Each group can ping at any time with status queries, escalations, new lead registrations, duplicate-detection flags. Multiply by 2,000+ groups and ops cost gets unsustainable.

Hiring more humans doesn't scale linearly — the cost per message becomes dominated by context-switching and routing, not by the actual answer.

The system we built

TRACE — Aminobots Assist under a GyanDhan brand — runs four agents in series. An Identifier classifies every message by intent and sentiment. A T0 rule layer pre-filters about half the traffic before any LLM is called — exact-match FAQ hits, intent-allow-list lookups, partner SLA template responses. Of what remains, a Support agent handles L1 resolution against the partner FAQ + CRM. An Expert agent handles edge cases — context that the L1 agent doesn't have. A HIL handoff routes the remainder to a human on the ops console.

Tiered LLM routing keeps cost down. About 80% of model calls go to the cheapest self-hosted OSS tier. 15% to a mid-tier model. 5% to a frontier model for edge cases. Average cost per query: $0.02.

What's underneath