RuleSense: Gen AI-powered Data Profiling for Regulatory Reporting

Table of Contents

• Introduction
• Demo
• Inspiration
• What It Does
• How We Built It
• Challenges We Faced
• How to Run
• Tech Stack
• Team
• Sample Dataset
• Regulatory Rules
• Submission Deliverables

---

Introduction

RuleSense is a Gen AI-powered compliance assistant designed to automate regulatory data profiling in the financial sector. It uses Gemini’s document processing capabilities and unsupervised machine learning models to extract rules from complex regulatory documents, validate transactional data, identify anomalies, assess risks, and recommend remediation actions.

---

Demo

• Live Demo: Link to Demo
• Video Walkthrough: Link to Video

Screenshots:

• RuleSense

• GenAI Rulebook Generator

• Regex Based Rulebooks

• Transactions Data Validator

• Anomaly Patterns Detector in Transactions
  

• Anomaly Detector
  

---

Inspiration

Financial institutions face a major challenge when translating regulatory reporting instructions into enforceable data rules. This manual and error-prone process often slows down compliance cycles. RuleSense was built to streamline this process by automatically interpreting regulations and applying intelligent validations.

---

What It Does

• Extracts data validation rules from regulatory PDFs using Gemini AI
• Parses extracted rules into structured JSON format with regex
• Validates uploaded transactional CSVs against these rules
• Flags validation issues and suggests remediation actions
• Detects anomalies in data using unsupervised learning
• Implements a dynamic risk scoring system
• Enables audit-friendly, explainable insights for compliance teams

sequenceDiagram
    %% Participants with icons
    actor User as User 
    participant UI as Flask HTML UI 
    participant Backend as Flask Backend (API Layer) 
    participant Gemini as Gemini AI API 
    participant RuleEngine as Rulebook Generator
    participant RegexParser as Regex Rule Parser 
    participant CSVValidator as CSV Validation Engine 
    participant AnomalyDetector as ML Anomaly Engine 
    participant JSONStore as Local JSON Rulebook

    %% PDF Upload & Rulebook Generation
    User->>UI: Upload regulatory PDF
    UI->>Backend: POST /generate-rulebook
    Backend->>Gemini: Send PDF content for processing
    Gemini-->>Backend: Extracted regulatory instructions
    Backend->>RuleEngine: Parse relevant rule sections
    RuleEngine->>RegexParser: Convert instructions to regex rules
    RegexParser-->>RuleEngine: Regex rules (field-wise)
    RuleEngine->>JSONStore: Save rulebook.json
    JSONStore-->>Backend: Confirmation
    Backend-->>UI: Display generated rulebook summary

    %% CSV Upload & Validation
    User->>UI: Upload CSV transactional file
    UI->>Backend: POST /validate-csv
    Backend->>JSONStore: Load rulebook.json
    Backend->>CSVValidator: Apply regex validations to CSV
    CSVValidator-->>Backend: Validation results (violations, matched rules)
    Backend-->>UI: Display validation results with rule references

    %% Anomaly Detection
    User->>UI: Click "Detect Anomalies"
    UI->>Backend: GET /detect-anomalies
    Backend->>CSVValidator: Preprocess numeric features
    CSVValidator->>AnomalyDetector: Send processed dataset
    AnomalyDetector-->>Backend: Anomaly labels and risk scores
    Backend-->>UI: Display anomaly summary and risk insights

    %% Remediation (Optional)
    Backend->>Backend: Generate remediation suggestions
    Backend-->>UI: Display automated recommendations

Loading

---

How We Built It

Regulatory Rulebook Extraction

• PDF parsing using Gemini AI’s document processing API
• Semantic chunking and cosine similarity for relevant rule matching
• Regex-based pattern extraction and rulebook generation (JSON format)

Validation Engine

• CSV upload interface for transactional datasets
• Field-by-field regex validation using the generated rulebook
• Detailed output: matched rule ID, violation description, suggestion

Anomaly Detection

• Uses Isolation Forest, DBSCAN, and LOF from scikit-learn
• Scales numeric features with StandardScaler
• Visualizes anomaly distribution using matplotlib

Frontend

• Built using flask templates to allow document and CSV uploads, validation display, and anomaly review

flowchart TD
  subgraph Client
    Browser["Web Browser (User)"]
  end

  subgraph UI["Flask Frontend"]
    HTMLUI["HTML Templates (Jinja2)"]
  end

  subgraph WebServer["Flask Backend"]
    Flask["Flask App"]
    API["API Routes"]
  end

  subgraph RulebookEngine["Rulebook Generation Engine"]
    Gemini["Gemini AI (Regulatory PDF Processor)"]
    RegexParser["Custom Regex Rule Parser"]
    RulebookJSON["Generated Rulebook (JSON)"]
  end

  subgraph Validator["CSV Validation Engine"]
    CSVUpload["CSV Upload & Parser"]
    RuleValidator["Regex Validator"]
    ValidationResults["Validation Report"]
  end

  subgraph AnomalyDetector["Anomaly Detection Engine"]
    Scaler["StandardScaler"]
    IFModel["Isolation Forest"]
    LOFModel["LOF / DBSCAN"]
    MLReport["Anomaly Report"]
  end

  subgraph Storage["Local Storage"]
    LocalFS["Filesystem (uploads, rulebooks, logs)"]
  end

  %% UI & Routing
  Browser --> HTMLUI
  HTMLUI --> Flask
  Flask --> API

  %% Rulebook Creation Flow
  API --> Gemini
  Gemini --> RegexParser
  RegexParser --> RulebookJSON
  RulebookJSON --> LocalFS

  %% CSV Validation Flow
  API --> CSVUpload
  CSVUpload --> RuleValidator
  RulebookJSON --> RuleValidator
  RuleValidator --> ValidationResults
  ValidationResults --> LocalFS

  %% Anomaly Detection Flow
  RuleValidator --> Scaler
  Scaler --> IFModel
  Scaler --> LOFModel
  IFModel --> MLReport
  LOFModel --> MLReport
  MLReport --> LocalFS

Loading

---

Challenges We Faced

• Parsing varied PDF structures and noisy formatting
• Extracting accurate regex patterns from natural language rules
• Balancing anomaly detection thresholds to avoid false positives
• Ensuring explanations are understandable by non-technical auditors
• Maintaining performance and scalability for large CSVs

---

How to Run

Clone the repository

git clone [https://github.com/your-username/rulesense.git](https://github.com/ewfx/gaidp-singularity-squad.git)
cd gaidp-singularity-squad.git

Set up environment variables

cp code/src/backend/.env.example code/src/backend/.env

Edit .env and update with your Gemini API key:

GOOGLE_API_KEY=your_api_key_here
FLASK_ENV=development
FLASK_DEBUG=1

Install Docker

• Follow official instructions to install Docker and Docker Compose:
• macOS/Linux: https://docs.docker.com/desktop/install/mac-install/
• Windows: https://docs.docker.com/desktop/install/windows-install/

Verify installation:

docker --version
docker compose version

Build and run the application

macOS/Linux

docker compose up --build

Windows

docker-compose up --build

Open the application

Once the container is running, go to:

http://localhost:5001

API Docs

For API Testing, go to:

http://localhost:5001/docs

Tech Stack

Languages & Frameworks

• Python 3.11
• Flask (Web framework)
• Jinja2 (HTML Templating)
• Scikit-learn (Machine Learning)
• Pandas, NumPy (Data processing)
• Matplotlib (Visualization)

AI & NLP

• Gemini AI API (Document understanding and rule extraction)

Validation Engine

• Custom Regex Parser (Rule interpretation and execution)

Storage

• Local filesystem for JSON rulebooks, uploads, and logs

DevOps & Tooling

• Docker (Development & deployment)
• dotenv (.env configuration)

Team

• Ayush Addhyayan – GitHub | LinkedIn
• Avirup Bhattacharyya – GitHub | LinkedIn
• Abhiram Bhimavarapu – GitHub | LinkedIn

Sample Dataset

This document describes the structure of each dataset used in the RuleSense project.

Transaction Validation Dataset

File: DatasetValidation.csv

This dataset is used to test rule validation logic, including regex-based validations extracted from regulatory instructions.

Schema

Column	Data Type	Description
customer_id	object	Customer id
committed_exposure_global	int64	Committed exposure global
utilized_exposure_global	int64	Utilized exposure global
line_reported_on_fr_y_9c	int64	Line reported on fr y 9c
credit_facility_currency	object	Credit facility currency
country	object	Country
origination_date	object	Origination date
obligor_internal_risk_rating	int64	Obligor internal risk rating

---

Anomaly Detection Dataset

File: DatasetAnomaly.csv

This dataset contains enriched features for training and evaluating unsupervised ML models like Isolation Forest, LOF, and DBSCAN.

Schema

Column	Data Type	Description
customer_id	object	Customer id
internal_id	object	Internal id
obligor_name	object	Obligor name
obligor_internal_risk_rating	object	Obligor internal risk rating
tin	object	Tin
country	object	Country
industry_code	object	Industry code
industry_code_type	object	Industry code type
internal_credit_facility_id	object	Internal credit facility id
origination_date	object	Origination date
maturity_date	object	Maturity date
credit_facility_type	object	Credit facility type
credit_facility_purpose	object	Credit facility purpose
committed_exposure_global	object	Committed exposure global
utilized_exposure_global	object	Utilized exposure global
credit_facility_currency	object	Credit facility currency
probability_of_default	object	Probability of default
loss_given_default	object	Loss given default
exposure_at_default	float64	Exposure at default
non_accrual_date	object	Non accrual date
days_principal_or_interest_past_due	float64	Days principal or interest past due
participation_flag	object	Participation flag
participation_interest	object	Participation interest
net_income_current	object	Net income current
total_assets_current	object	Total assets current
long_term_debt	float64	Long term debt
current_assets_current	object	Current assets current
current_liabilities_current	object	Current liabilities current
guarantor_flag	object	Guarantor flag
guarantor_internal_id	object	Guarantor internal id
guarantor_name	object	Guarantor name
guarantor_tin	object	Guarantor tin
guarantor_internal_risk_rating	object	Guarantor internal risk rating
entity_internal_id	object	Entity internal id
entity_name	object	Entity name
entity_internal_risk_rating	object	Entity internal risk rating
date_of_financials	object	Date of financials
date_of_last_audit	object	Date of last audit
cash_marketable_securities	object	Cash marketable securities
accounts_receivable_current	object	Accounts receivable current
inventory_current	object	Inventory current
fixed_assets	object	Fixed assets
total_assets_prior_year	object	Total assets prior year
current_liabilities_prior_year	object	Current liabilities prior year
short_term_debt	object	Short term debt
capital_expenditures	object	Capital expenditures
asc_326_20	object	Asc 326 20
special_purpose_entity_flag	object	Special purpose entity flag
other_credit_facility_type_description	object	Other credit facility type description
original_internal_id	object	Original internal id
stock_exchange	object	Stock exchange
ticker_symbol	object	Ticker symbol
cusip	int64	Cusip
original_internal_credit_facility_id	object	Original internal credit facility id
line_reported_on_fr_y_9c	int64	Line reported on fr y 9c
line_of_business	object	Line of business
cumulative_charge_offs	int64	Cumulative charge offs
days_principal_or_interest_past_due.1	int64	Days principal or interest past due.1
lien_position	int64	Lien position
security_type	int64	Security type
interest_rate_variability	int64	Interest rate variability
interest_rate_index	int64	Interest rate index
interest_rate_spread	float64	Interest rate spread
interest_rate_ceiling	float64	Interest rate ceiling
interest_rate_floor	float64	Interest rate floor
interest_income_tax_status	int64	Interest income tax status
guarantor_internal_risk_rating.1	object	Guarantor internal risk rating.1
entity_internal_risk_rating.1	object	Entity internal risk rating.1
net_sales_current	int64	Net sales current
net_sales_prior_year	int64	Net sales prior year
operating_income	int64	Operating income
depreciation_amortization	int64	Depreciation amortization
interest_expense	int64	Interest expense
accounts_receivable_prior_year	int64	Accounts receivable prior year
inventory_prior_year	int64	Inventory prior year
current_assets_prior_year	int64	Current assets prior year
tangible_assets	int64	Tangible assets
total_assets_current.1	int64	Total assets current.1
accounts_payable_current	int64	Accounts payable current
accounts_payable_prior_year	int64	Accounts payable prior year
current_maturities_of_long_term_debt	int64	Current maturities of long term debt
long_term_debt.1	int64	Long term debt.1
minority_interest	int64	Minority interest
total_liabilities	int64	Total liabilities
retained_earnings	int64	Retained earnings
lower_of_cost_or_market_flag	int64	Lower of cost or market flag
snc_internal_credit_id	object	Snc internal credit id
renewal_date	object	Renewal date
prepayment_penalty_flag	int64	Prepayment penalty flag
entity_industry_code	int64	Entity industry code
leveraged_loan_flag	int64	Leveraged loan flag
disposition_flag	int64	Disposition flag
disposition_schedule_shift	object	Disposition schedule shift
syndicated_loan_flag	int64	Syndicated loan flag
target_hold	float64	Target hold
pcd_noncredit_discount	int64	Pcd noncredit discount
committed_exposure_global_par_value	int64	Committed exposure global par value
utilized_exposure_global_par_value	int64	Utilized exposure global par value
committed_exposure_global_fair_value	int64	Committed exposure global fair value
utilized_exposure_global_fair_value	int64	Utilized exposure global fair value
obligor_lei	object	Obligor lei
primary_source_of_repayment_lei	object	Primary source of repayment lei

Regulatory Rules

RuleSense leverages regulatory instructions extracted from official reporting documents such as the FR Y-14Q.

The extracted rules are used for validating transactional data in line with compliance expectations. These instructions are manually and LLM-assisted parsed from original documentation.

📄 View the extracted regulatory instructions (PDF format):
Download Rulebook.pdf

Submission Deliverables

• ✅ Flask backend with PDF parsing, rule generation, validation, and anomaly detection
• ✅ Jinja2 HTML UI for CSV upload and result display
• ✅ Gemini-powered rule extraction engine
• ✅ Regex-based rule validation engine
• ✅ Isolation Forest and LOF for anomaly detection
• ✅ Three datasets:
  • Validation
  • Anomaly Detection
  • Clean Baseline
• ✅ Dataset schema
• ✅ Rulebook Pdf
• ✅ Sequence & tech stack diagrams: Mermaid-based
• ✅ .env.example and Docker setup
• ✅ Documentation and project README
• ✅ Demo video