DEV Community

pybankers
pybankers

Posted on Edited on Originally published at pybankers.com

CAG Built Its Own LLM: Inside India's Sovereign Audit AI

CAG Built Its Own LLM: Inside India's Sovereign Audit AI

PyBankers — CAG built its own LLM

On 16 November 2026, the Comptroller and Auditor General of India (CAG) unveils LLM 2.0 — an in-house large language model developed by IIT Madras and trained on roughly 65,000 of the organisation's own internal inspection reports, plus its auditing standards and guidelines.

This is an unusual case: a national constitutional audit authority shipping its own domain-specific model instead of renting a general-purpose one from a commercial API.

What the model actually is

Property Detail
Developer IIT Madras, built for CAG
Training corpus ~65,000 internal inspection reports, auditing standards, guidelines
Architecture Three streams — compliance audit, financial audit, works audit
Launch 16 November 2026, at Audit Diwas
Public access None announced — no API, no pricing, no timeline for industry use

The three streams

Each audit type works on materially different evidence, so the model is partitioned rather than generalised:

  • Compliance audit — rule-based checks against delegated legislation and spending rules
  • Financial audit — financial statements, auditor's reports, board bills
  • Works audit — infrastructure and construction expenditure

Auditors upload their source documents and the model handles the mechanical layer of preparation.

The human boundary is explicit

Deputy CAG Krishnan Sangaran Subramanian framed the role clearly: the LLM performs numerical calculations and surfaces patterns. The judgement and the sign-off stay with the auditor. It is an assistant, not an autonomous auditor, and there is no automated audit decision path.

A 14-month build-up, not a surprise

  • April 2025 — CAG's Artificial Intelligence Strategy Framework, authored by CTO S. Ramann
  • September 2025 — the CAG-LLM named publicly in a press release as an initiative under development
  • August 2026 — Certification on Audit of Digital Systems with IIT Madras; a nine-month internal course on data analytics, AI and cybersecurity for CAG officers
  • 16 November 2026 — LLM 2.0 launch

Audit planning is changing alongside it: 2026-27 plans move away from unit-centric selection toward choosing audits by data patterns across departments.

Why sovereignty, structurally

For a constitutional body auditing government expenditure, uploading inspection reports to a third-party API is a data-sovereignty question rather than a procurement preference. Building in-house resolves it by construction.

What has not been published

No parameter count, base model, architecture or context window has been disclosed. No accuracy, recall or hallucination rate has been published by either CAG or IIT Madras. Any performance claim here would be vendor-reported until independently reproduced. There is no public access route.

The pattern worth noting

Three elements transfer to any organisation that holds a large corpus of its own domain knowledge:

  1. Own corpus over rented generalisation — a compact model trained on your actual document type behaves differently from a frontier API given examples.
  2. Partition by workflow, not by user type — three narrow streams beat one general model when evaluation criteria differ per task.
  3. Strategy before model — the strategy framework preceded the model by fourteen months.

The same idea at a smaller scale: our offline GST invoice engine distils a document transformer until it runs on a shop's own machine, measuring 95.95% entity F1 across 136 held-out invoices — because it was trained on the actual document type rather than generic web text.

Sources

Full analysis with the scale comparison is on PyBankers:
https://pybankers.com/blog/cag-in-house-llm-2-0-iot-madras-sovereign-audit-ai

Top comments (0)