🛡️ SEBI Compliance MCP System (Working End-to-End Demo)

A deterministic regulatory compliance evaluation system built for SEBI (Research Analyst) Regulations, 2014.

This project exposes regulatory compliance rules as Model Context Protocol (MCP) tools served by a FastMCP (streamable-http) server. The rules are queried deterministically from a structured JSON file (data/rules.json) without any database, RAG pipeline, or vector store. A LangGraph React Agent powered by Groq (ChatGroq) acts as the MCP client, orchestrating tool calls based on natural language queries and returning verifiable, deterministic audit trails.

🏗️ Architecture & Core Principles

+-------------------------------------------------------------------------------+
|                        Streamlit Frontend (Port 8501)                         |
|   - Chat Interface & Expandable Deterministic Tool Audit Trails               |
|   - Preset Scenario Buttons for One-Click Live Demos                          |
+-------------------------------------------------------------------------------+
                                       |
                             POST /chat (JSON)
                                       v
+-------------------------------------------------------------------------------+
|                       FastAPI Backend (Port 8001)                             |
|   - MultiServerMCPClient (langchain-mcp-adapters)                             |
|   - LangGraph React Agent (langgraph.prebuilt.create_react_agent)             |
|   - Groq LLM (ChatGroq) orchestration & tool selection logic                  |
+-------------------------------------------------------------------------------+
                                       |
                   MCP Streamable HTTP Transport (Port 8000)
                                       v
+-------------------------------------------------------------------------------+
|                      FastMCP Server (`mcp_server/server.py`)                  |
|   - Tools: check_compliance, get_applicable_rules, get_audit_log              |
|   - Deterministic evaluation using safe operator map (NO eval())              |
+-------------------------------------------------------------------------------+
                                 /            \
                  Reads Rules   /              \  Writes Audit Log
                               v                v
          +------------------------+      +---------------------------+
          |    data/rules.json     |      |    audit/audit_log.json   |
          |  (14 Verified Rules)   |      |  (Persistent Log Array)   |
          +------------------------+      +---------------------------+

🎯 Why Deterministic MCP?

Compliance verdicts (pass, fail, needs_review) must come exclusively from deterministic code evaluating structured JSON rules — NEVER from an LLM judging raw text or hallucinating regulatory limits.

• The LLM's only job is:

  1. Deciding which MCP tool(s) to call based on the user's natural-language question (check_compliance, get_applicable_rules, get_audit_log).

  2. Turning the structured tool output into a readable response while citing exact SEBI regulations and displaying the unique audit_id.

• No eval(): Numeric and boolean rule conditions are evaluated strictly via safe operator mappings (operator.ge, operator.le, operator.eq, operator.contains).

• Qualitative Rules: Clauses that require human judgment or documentary verification (e.g. conflict of interest policies or research rationale reasonableness) use operator: "manual_review" and deterministically output needs_review.

Related MCP server: @run-iq/mcp-server

📂 Project Structure

sebi-compliance-mcp/
├── .env.example              # Example environment configuration
├── .env                      # Local environment variables (GROQ_API_KEY)
├── requirements.txt          # Pinned Python package dependencies
├── README.md                 # Project documentation & run guide
├── data/
│   └── rules.json            # 14 hand-authored SEBI Research Analyst rules
├── mcp_server/
│   └── server.py             # FastMCP server exposing compliance evaluation tools
├── backend/
│   └── app.py                # FastAPI backend + LangGraph React Agent + MCP Client
├── frontend/
│   └── streamlit_app.py      # Streamlit interactive chat UI
└── audit/
    └── audit_log.json        # Persistent JSON audit log (generated at runtime)

🚀 Setup & Installation

1. Prerequisites

• Python 3.11+ installed locally.

• A Groq API Key (GROQ_API_KEY). You can get one for free at console.groq.com.

2. Create Virtual Environment & Install Dependencies

# Create and activate virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

3. Configure Environment Variables

Copy .env.example to .env and add your Groq API key:

cp .env.example .env

In .env:

GROQ_API_KEY=your_actual_groq_api_key_here
GROQ_MODEL=llama-3.3-70b-versatile

▶️ Running the System (Exact Run Order)

Open three separate terminal windows/tabs (make sure your virtual environment venv is activated in all three):

Terminal 1: Start FastMCP Server (Port 8000)

python mcp_server/server.py

You should see output indicating the server is running on http://0.0.0.0:8000 (streamable-http).

Terminal 2: Start FastAPI Backend (Port 8001)

uvicorn backend.app:app --reload --port 8001

On startup, the backend connects to http://localhost:8000/mcp, initializes the MultiServerMCPClient, and registers tools with the LangGraph React Agent.

Terminal 3: Start Streamlit Frontend (Port 8501)

streamlit run frontend/streamlit_app.py

Your browser will automatically open at http://localhost:8501 showing the interactive chat UI.

🛠️ MCP Tools Reference (mcp_server/server.py)

1. check_compliance(entity_type: str, scenario: dict) -> dict

   • Loads rules matching entity_type (individual_RA, partnership_RA, non_individual_RA).

   • Evaluates each scenario attribute against rules.json.

   • Computes overall_status: "fail" if any rule failed, else "needs_review" if any rule requires manual verification or scenario fields are missing, else "pass".

   • Generates a unique audit_id (e.g. q-2026-07-12T15:30:00Z-a1b2c3) and saves the exact execution record to audit/audit_log.json.

2. get_applicable_rules(entity_type: str) -> list

   • Returns all verified SEBI rules and regulatory citations applicable to the specified entity type without evaluating a scenario.

3. get_audit_log(audit_id: str) -> dict

   • Retrieves the persistent log entry from audit/audit_log.json by its unique audit_id.

📋 Rule Schema (data/rules.json)

Each rule adheres strictly to the following schema:

{
  "rule_id": "SEBI-RA-2014-3.2-a",
  "regulation": "SEBI (Research Analysts) Regulations, 2014",
  "clause_ref": "Regulation 3(2)(a)",
  "entity_type": "individual_RA",
  "condition_type": "net_worth_min",
  "operator": ">=",
  "value": 100000,
  "unit": "INR",
  "effective_date": "2014-09-01",
  "effective_until": null,
  "source_citation": "SEBI (Research Analysts) Regulations, 2014, Reg 3(2)(a)",
  "status": "verified"
}

Covered Conditions & Operators

• Numeric Limits: net_worth_min, experience_years_min, trading_window_lockin_days_before/after, record_keeping_years_min, holding_in_subject_company_max_pct (>=, <=, ==)

• Categorical & Set Matches: qualification_req (in array of allowed qualifications)

• Boolean Requirements: nism_certification_active, compliance_officer_appointed, compensation_from_merchant_banking (==)

• Qualitative Standards (operator: "manual_review"): conflict_of_interest_policy_quality, research_rationale_basis_adequacy (returns needs_review with regulatory targets)

💡 Example Demo Queries

Try these out in the Streamlit UI or by clicking the preset sidebar buttons:

• Individual RA (Should Fail):

  "Check compliance for an individual_RA with net_worth_min of 80000 INR, qualification_req of post_graduate_degree, nism_certification_active true, experience_years_min of 5, trading_window_lockin_days_before 30, trading_window_lockin_days_after 5, compensation_from_merchant_banking false, holding_in_subject_company_max_pct 0.5, and record_keeping_years_min 5."

• Non-Individual RA (Should Pass / Review):

  "Verify compliance for a non_individual_RA with net_worth_min 3000000, compliance_officer_appointed true, record_keeping_years_min 6, conflict_of_interest_policy_quality documented_and_effectively_enforced, and compensation_from_merchant_banking false."

• Missing Information (Should Trigger Clarification / Needs Review):

  "Check compliance for an individual RA with ₹200,000 net worth."