Gujarat Police Deploys NARIT-AI To Strengthen NDPS Prosecutions

The Gujarat Police rolled out NARIT-AI (Narcotics Analysis & RAG-based Investigation Tool) to tighten prosecutions under the NDPS Act and reduce acquittals caused by procedural lapses. The system was conceived by Abhay Soni and developed by the Western Railway Police in collaboration with Gradiante Creative Services. It was built and field-tested in three months, with field trials conducted during the final month of development, and is designed to convert uploaded FIRs into legally grounded investigative plans, evidence checklists, draft chargesheets, and court summaries.

NARIT-AI uses RAG (retrieval-augmented generation) architecture operating over a curated, closed knowledge base that includes bare acts, government circulars, and thousands of Supreme Court and High Court judgments. Because it does not query the open internet, the system is configured to minimize hallucinations and supply traceable citations.

• •The platform accepts an FIR upload and returns a case-specific investigative roadmap with timelines and mandatory actions to preserve admissible evidence.
• •Outputs include an evidence checklist tailored to the narcotic type, a draft chargesheet, a prosecutorial summary intended for magistrates, and predicted defense arguments with suggested rebuttals.
• •Security controls are enforcement-grade, restricted to verified police users, and the knowledge base is designed for dynamic updates when laws or circulars change.

Prosecutors in Gujarat and elsewhere have struggled with NDPS cases because minor procedural errors frequently trigger acquittals; conviction rates in some reports fell to roughly 25%. NARIT-AI operationalizes a practical tradeoff: reduce model creativity by anchoring generation to a vetted legal corpus, and in return produce auditable, jurisdiction-specific guidance that nonlegal investigators can act on. For the ML community, it is a field example of how RAG can be applied to high-stakes, compliance-heavy domains where provenance, updateability, and access control matter more than raw generative flexibility.

The tool promises faster, more consistent adherence to NDPS procedures, potentially increasing successful prosecutions. However, risks remain: overreliance on automated guidance could entrench procedural heuristics that miss novel evidentiary scenarios. There are also governance and accountability questions around who vets the system's legal interpretations, and how defense counsel can challenge algorithm-driven prosecutorial recommendations. Transparency mechanisms, audit logs, and human-in-the-loop review will be decisive for legal defensibility.

Monitor adoption and measurable outcomes, specifically changes to conviction rates and appellate reversals linked to NARIT-AI usage. Also watch for published technical audits, the dataset governance model, and whether similar RAG-based tools proliferate across other Indian states or case types.

NARIT-AI is a clear blueprint for domain-locked RAG deployments: curate high-quality legal sources, enforce strict access and update controls, and present outputs as prescriptive checklists plus human-reviewable documents rather than as final legal opinions. For ML engineers working in regulated domains, the implementation choices here, closed sandbox, auditable citations, and role-based access, are instructive design patterns.