NITI Aayog selected six states—Rajasthan, Madhya Pradesh, Odisha, Assam, Telangana, and Gujarat—to pilot a vernacular large-language-model layer that sorts block-level complaints before they reach district collectors, according to a project brief shared with state IT secretaries on Friday. The system does not decide outcomes; it tags grievances by scheme, urgency, and language, then routes them into existing CM dashboard workflows.
What the model is allowed to do
Complaints arrive today as unstructured voice notes, WhatsApp forwards, and handwritten scans. Human clerks spend hours retyping them into English templates. The pilot model—hosted on state data-centre hardware segregated from public cloud APIs—will transcribe Hindi, Odia, Telugu, and Gujarati inputs, then suggest a category code from the Centre’s standard grievance taxonomy. A human operator must confirm every tag before the ticket enters the collector’s queue; auto-closure is explicitly banned in the memorandum.
Training data comes from redacted 2024–2025 tickets supplied by the states, roughly 1.2 million records after personal identifiers are stripped. NITI Aayog’s technical partner list includes two Indian startups and a public-sector unit already running air-gapped inference for tax analytics. No foreign foundation model may be called at runtime; weights must be deployable offline.
Why block level matters
Most scheme failures surface at the gram panchayat or urban ward office long before they trigger a headline in the state capital. MGNREGA wage delays, PMAY installment holds, and ration shop stock-outs often share wording patterns that confuse manual triage. The LLM layer is tuned to spot those patterns— for example, a Telugu voice note mentioning “job card slip” and “14 days” should map to a labour department SLA breach rather than a generic “scheme complaint.”
Risks the memo acknowledges
Bias audits are scheduled every ninety days, with mandatory rollback if false routing exceeds 8 percent on any single scheme code. States must publish aggregate accuracy metrics but not raw citizen messages. Privacy advocates cited in policy press want independent access logs; the brief promises quarterly reviews by state information commissions. Funding is ₹48 crore across eighteen months, split between GPU procurement and district training workshops.
Rollout calendar
Rajasthan and Telangana go live in November on a single district each; the other four states join in January after localization for Assamese and Odia dialect variants. If the pilot cuts median resolution time without increasing appeals, NITI Aayog will propose a national API specification for grievance triage—not a mandate, but a template other states could adopt without buying the same vendor stack.
Union vs state IT stacks
Several states already run separate grievance apps that do not interoperate; the pilot’s API layer is meant to sit above those silos rather than replace them on day one. District collectors will still see one merged queue per morning, colour-coded by SLA risk. Union ministries asked for read-only feeds into their dashboards so scheme-specific escalations do not duplicate work at the Centre.