An Agentic AI Layer for India's Last-Mile Public Delivery Gap.

India operates two of the world's largest public service delivery networks: National Tuberculosis Elimination Programme (NTEP) serving 1.4B people, and agricultural extension supporting over 140 million smallholder farming households. Both face severe human capacity bottlenecks.

The operational breakdown is not policy availability or digital registry infrastructure — it is frontline expert bandwidth. Key challenges across public health and agriculture include:

• Tuberculosis Diagnostic Delays: India accounts for 25% of global TB incidence and ~33% of drug-resistant cases. Upstream diagnostic delays occur because initial screening relies on lab-bound testing concentrated in urban centers.
• Agricultural Extension Shortages: Recommended extension officer ratio is 1:750, yet India operates far below that nationally. Farmers lose 10-30% of crops (₹1.7–3.42 lakh crore annually) to pests due to late or incorrect diagnostic advice.
• Connectivity & Digital Literacy Barriers: Remote rural centers lack continuous high-speed internet, rendering cloud-dependent LLM chatbots unreliable for field workers and farmers.
• Generic Chatbot Unreliability: General-purpose AI chatbots produce confident wrong answers, lack bounded domain protocols, and cannot perform offline audio/image diagnostic classification.

Instead of building another general chatbot, we deploy a narrow, accountable Agentic AI decision-support layer operating directly on low-cost Android hardware.

The agent executes on-device computer vision and acoustic classification offline, provides source-grounded triage, and explicitly escalates uncertain cases to human specialists:

• TB Screening & Referral Agent: Processes cough audio + symptom checklists offline via Android phones. High-risk or low-confidence results trigger instant referral to sputum/NAAT molecular testing in national reporting channels (Ni-kshay).
• Crop Pest & Disease Advisory Agent: Analyzes leaf photos and voice descriptions in local dialects. High-confidence matches return government-approved chemical treatments; low-confidence matches auto-escalate to human agronomists.
• Uncertainty Escalation by Default: Built-in confidence thresholds ensure the system defaults to specialist referral when uncertain, preventing dangerous misdiagnoses or improper pesticide usage.
• Voice-First Vernacular Interaction: Supports code-switched regional dialects and voice interaction, eliminating text typing requirements for low-literacy users.

Analysis of key operational risks and engineering safeguards for field deployment: