Prahari

What it is

The product

Prahari (Hindi: sentinel) is an agentic perception-intelligence platform for broadcast media. One agent per channel captures the feed, transcribes it in-language and reads the on-screen text; a resolution agent matches every mention against a configurable graph of principals, institutions, schemes and opponents; a classification agent scores stance relative to the principal, framing, speaker role, prominence and tone, with the clip attached as evidence; a narrative agent correlates across channels and languages to separate an organic frame from a coordinated one; and an escalation agent decides what deserves a human — a decision card on WhatsApp within two minutes, or a line in the autonomous daily digest. A longitudinal perception graph answers questions conversationally, and a guardian agent audits coverage completeness and classification neutrality.

The problem

What we set out to solve

A public institution cannot watch television at the scale and speed that narrative moves. One state has dozens of channels across several languages running around the clock, and a hostile frame can form on four of them inside ninety minutes. Human monitoring catches it hours late, cannot correlate across channels or languages, cannot tell organic sentiment from a coordinated push, and cannot promise it saw everything. The incumbent vendors return keyword clip-dumps: they move the haystack rather than remove it, and offer no completeness guarantee.

The build

How it actually works

The pipeline is source-agnostic after entity resolution, so broadcast, live-stream, press and social all normalise to the same perception event. Indian-language ASR runs through a tiered provider chain with a heuristic fallback, and the classifier degrades visibly rather than silently when the model is unreachable. Roughly ninety-nine percent of broadcast segments are shed before any model call, which is what keeps the fleet affordable at scale. Every channel keeps a coverage ledger so gaps are logged, not hidden. The first deployment runs against a state government’s broadcast environment, Kannada-first; the second points the same engine at a professional sports franchise’s fan base from public signals alone. Pilots run on a single in-region VM with data residency in India, and decision cards are delivered through Breez.

Capabilities

• Zero-watch capture across every configured channel and language
• Cross-script, phonetic entity resolution with namesake suppression
• Principal-relative stance, framing, speaker role and prominence per mention
• Coordinated-versus-organic narrative attribution across channels
• Tiered escalation with the evidence clip, and an autonomous daily digest
• Per-channel and per-anchor treatment scorecards over time
• Conversational questions over a longitudinal perception graph
• Coverage-completeness and neutrality audit by a guardian agent