KKissan Ki Pehchan
Architecture

End-to-End Case Workflow

Stateful live-call workflow from room join through adaptive evidence, research, answer and human handoff.

BlueprintVersion 0.25 Aug 2026

Workflow

START CASE
  ↓ consent + room token
JOIN LIVE AUDIO/VIDEO ROOM
  ↓ streaming STT + low-rate visual sampling
FRAME CASE / IDENTIFY INTENT
  ↓
COLLECT CONTEXT
  crop · location · stage · treatments · weather · soil
  ↓
ACTIVE EVIDENCE LOOP
  ask one question OR issue one camera objective
  → quality-check selected frame/still
  → update hypotheses and missing information
  → repeat only while information value justifies it
  ↓
LOCAL RETRIEVAL + OPTIONAL WEB RESEARCH
  ↓
DIFFERENTIAL ANALYSIS + SELF-CHALLENGE
  ↓
EXACT POLICY VERIFICATION
  ↓
RISK / SUFFICIENCY GATE
  ├── answer in Urdu
  ├── ask/re-capture
  ├── officer joins live
  ├── warm handoff
  └── persist for asynchronous review
  ↓
AUDIT + FOLLOW-UP + END ROOM

Live state

Room connectivity and case reasoning are related but separate state machines. The case can continue through reconnects, degraded media and post-call research. Every accepted turn and selected frame belongs to an immutable reasoning snapshot.

Long operations

Deep research, expensive video processing and human review may continue after the live call. The farmer receives immediate acknowledgement, a case identifier and notification status. New corrections or media create a new immutable snapshot without deleting history.

Stopping and handoff

The agent stops visual requests when evidence is sufficient for the risk, further capture has low expected value, the farmer cannot provide the view, or a human/laboratory pathway is more appropriate. It must not prolong a call merely to increase confidence numerically.