Agentic RBI Compliance Intelligence System

RBI Regulatory Change | Evidence-Grounded | Role-Aware | Auditable

The Problem

RBI issues regulatory publications continuously. A compliance officer, a finance leader, a credit leader, and an operations team may all need different answers from the same publication: what changed, what it replaces, whether it applies to their function, and what needs to happen next.

The problem is not access to RBI publications. The problem is the distance between a published regulatory document and an actionable operational brief, with enough evidence for a human reviewer to trust, challenge, or reject the interpretation.

This project explores that gap. It is a prototype system for turning public RBI material into evidence-grounded, role-aware review artifacts with source traceability, validation signals, review queues, local persistence, and notification support.

The aim is to ensure that every relevant regulatory publication can reach the right reviewer with a validated, role-specific brief and enough supporting context to understand what changed, why it matters, and what was recorded.

The Design Question

How do you build a system a compliance reviewer can inspect and trust, not just one that summarizes?

The project treats regulatory interpretation as a provenance problem. Generated outputs should remain connected to source material, uncertainty should stay visible, and durable acceptance should remain separate from automated generation. A human reviewer remains responsible for deciding whether any output is correct, relevant, or actionable.

What It Does

The app helps a reviewer work through the lifecycle of RBI regulatory-change intelligence:

RBI source
  -> source ingestion
  -> relevance filtering
  -> change interpretation
  -> evidence validation
  -> human review
  -> local persistence
  -> notification support

In practice, it provides:

• a pipeline UI for running RBI source processing manually or on a schedule;
• source review queues for skipped, deferred, or partially classified publications;
• review queues for source records, generated interpretations, and related-document signals;
• guided local corpus preparation for grounding retrieval in public RBI material;
• a query and exploration surface for inspecting the local corpus, retrieved context, review outputs, and source-backed findings;
• configurable institution and role profiles that shape relevance filtering, impact framing, audience hypotheses, notification decisions, and suppression rationale;
• local persistence for reviewable source, output, and execution records;
• notification support with generated role-oriented summaries that remain reviewable.

What It Is Not

This project is not:

• a full compliance management platform with obligation libraries, task workflows, attestations, controls testing, or audit management;
• a regulatory reporting engine;
• a legal advice product;
• a production GRC system;
• a general-purpose document Q&A bot.

It is decision-support software for research and review workflows. Outputs must be checked against official RBI publications and professional judgment before use.

Architecture

See ARCHITECTURE_OVERVIEW.md for the public-safe architecture summary.

The short version: the system uses an 8-node LangGraph workflow whose agentic stages classify new RBI sources, retrieve prior regulatory context, reason over amendments and supersessions, validate evidence, map operational impact, prepare reviewable outputs, and support role-specific notification decisions. Deterministic code handles parsing, normalization, storage boundaries, review gates, and auditability.

The local corpus follows an LLM-wiki style pattern with retrieval-augmented workflows, including embedding-backed lookup where appropriate. A separate corpus-editing agent is planned for controlled maintenance of that local corpus, with review gates before durable edits.

Institution and role profiles shape how the workflow interprets the same RBI source for different regulated-entity scopes and reviewer audiences. The public repo describes this capability without publishing private inventories, routing rules, role mappings, or notification policies.

Local Setup

uv sync
cp .env.example .env
uv run python app_pipeline.py

The local environment file must be populated with the model provider credentials required by the configured pipeline. Private prompts, production configuration, local data stores, and generated corpus artifacts are not included in the public repository.

Local Corpus

The project includes guided workflows for preparing a local RBI reference corpus from public regulatory material, so retrieval and interpretation can be grounded in source evidence.

The public repository describes this capability at a high level. Private seed procedures, generated review artifacts, local corpus files, retrieval indexes, repair utilities, and operational run outputs are intentionally excluded from source control.

Evaluation

The repository includes a public-safe evaluation overview and harness structure. Private fixtures, protected cohorts, detailed rubrics, generated run artifacts, and scoring history are intentionally excluded.

See:

• eval/EVALUATION.md

Repository Boundary

The public source release is intended to include:

• application and pipeline source code;
• public-safe documentation;
• selected eval harness code and public-safe corpus scaffolding;
• tests and sample fixtures that are safe to publish;
• example configuration templates.

It intentionally excludes:

• secrets and environment files;
• private prompts and production configuration;
• local data stores and retrieval indexes;
• generated wiki corpus files and PDFs;
• private eval fixtures and results;
• internal transition notes and internal review notes;
• generated seed workbooks, logs, cache files, and run artifacts.

Disclaimer

This project does not provide legal, regulatory, audit, or compliance advice. Users must verify all outputs against official RBI publications and their own professional judgment before taking action.

See DISCLAIMER.md.