AuditGPT

NSE Forensic Intelligence Engine
Quantitative fraud detection for Indian equity markets

Built for IAR Udaan Hackathon 2026 · Day 3 · Problem #01 · Solo · 30 hours

---

Input a company name → AuditGPT fetches 10 years of NSE financial data → runs four forensic models simultaneously → outputs a composite fraud risk score (0–100), an anomaly heatmap, a red flag timeline, peer sector comparison, sentiment trend, and a streaming LLM forensic narrative. Every signal is benchmarked against industry peers, not just absolute thresholds.

---

NSE Forensic Intelligence Engine — quantitative fraud detection for Indian equity markets, powered by Beneish M-Score, Altman Z-Score, industry-adjusted anomaly detection, and Gemini LLM narrative synthesis.

Pages

Route	Page	Description
/	Landing	Marketing page with live demo strip and model overview
/radar	Fraud Radar	Sector grid sorted by risk · search · company drill-down panel
/report/:id	Forensic Report	Full dashboard for a single company
/critical	Critical Section	NSE-wide threat matrix heatmap + contagion splash zone
/satyam	Satyam Case Study	Year-by-year reconstruction of India's largest corporate fraud (2000–2009)

---

Tech Stack

Layer	Technology
Frontend	React 18 + Vite
Backend	FastAPI (Python 3.11+) · Uvicorn
Hosting	Vercel (frontend) · Railway (backend)
Quant Engine	pandas · numpy · scipy
LLM	Google Gemini 1.5 Flash · SSE streaming
Sentiment	VADER
Charts	Recharts · custom SVG
Data	Pre-cached JSON (Screener.in format) · no database
Fonts	JetBrains Mono · Space Grotesk

Design system: Bloomberg Terminal dark — #070b12 background · #00ff88 green · #ffb020 amber · #ff4455 red · #00d4ff cyan.

---

Detection Models

Composite Score Weights

Model	Weight	What It Detects	Normalization
Beneish M-Score	35%	Earnings manipulation via 8 accrual variables	min(max((M+3)/5×100, 0), 100)
Altman Z-Score	30%	Financial distress / bankruptcy risk	min(max((4−Z)/4×100, 0), 100)
Industry-Adjusted Z	25%	Peer-relative ratio outliers (12 ratios)	min((avg_abs_z/3)×100, 100)
Trend Breaks	10%	Structural breaks in financial time series	(break_count/12)×100

Risk thresholds: 0–25 Low · 26–50 Medium · 51–75 High · 76–100 Critical

Beneish M-Score Variables

DSRI · SGI · GMI · AQI · DEPI · SGAI · TATA · LVGI — all 8 variables decomposed individually, charted over time, and compared against the −1.78 manipulation threshold.

Altman Z-Score Zones

Zone	Range
Safe	Z > 2.99
Grey	1.81 < Z < 2.99
Distress	Z < 1.81

---

Backtested Accuracy

Model	Metric	Result
Beneish M-Score	Detection rate	76%
Altman Z-Score	2-year prediction accuracy	82%
Industry Z-Score	Anomaly detection accuracy	71%
Ensemble (all 4)	Combined accuracy	92%

---

Getting Started

Prerequisites

• Python 3.11+
• Node.js 18+
• Google Gemini API key (free tier works)

Install

git clone https://github.com/your-username/auditgpt
cd auditgpt

# Backend
pip install -r backend/requirements.txt

# Frontend
cd frontend && npm install

Environment Variables

# backend/.env
GEMINI_API_KEY=your_key_here

# frontend/.env.local
VITE_API_URL=http://localhost:8000/api

Add Company Data

Drop Screener.in-format JSON files into backend/data/companies/. Each file should be named {COMPANY_ID}.json.

Precompute Scores

Walks all company JSONs, runs the full quant pipeline, builds peer groups by sector, and writes reports/{id}.json + sector_summary.json.

python scripts/precompute.py

Run

# Terminal 1 — Backend
uvicorn backend.main:app --reload --port 8000

# Terminal 2 — Frontend
cd frontend && npm run dev

Open http://localhost:5173

---

API Reference

GET  /api/sectors                  → Sector summary sorted by avg risk score desc
GET  /api/sectors/{sector_name}    → All companies in sector sorted by risk desc
GET  /api/search?q={query}         → Fuzzy search, returns top 10 matches
GET  /api/report/{company_id}      → Full ForensicReport (scores + peers + financials + sentiment)
GET  /api/stream/{company_id}      → SSE stream of Gemini LLM narrative

Company JSON shape (Screener.in format)

{
  "company_id": "TCS",
  "company_name": "Tata Consultancy Services",
  "sector": "IT Services",
  "market_cap": 868235.0,
  "profit_loss": {
    "Sales+": { "Mar 2014": 81809, "Mar 2015": 94648 },
    "Net Profit+": {}
  },
  "balance_sheet": {
    "Total Assets": {},
    "Reserves": {},
    "Borrowings+": {}
  },
  "cash_flow": {
    "Cash from Operating Activity+": {}
  },
  "ratios": {
    "Debtor Days": {},
    "Working Capital Days": {}
  }
}

Years are "Mar YYYY" format. TTM and Sep/Jun quarter keys are filtered out by the engine.

ForensicReport response shape

{
  "company_id": "TCS",
  "company_name": "Tata Consultancy Services",
  "sector": "IT Services",
  "composite_score": 11.4,
  "risk_level": "LOW",
  "risk_reasoning": "...",
  "breakdown": {
    "beneish_normalized": 8.2,
    "altman_normalized": 12.1,
    "industry_z_normalized": 14.0,
    "trend_break_normalized": 0.0
  },
  "beneish": { "m_score": -2.91, "manipulation_likely": false, "components": {} },
  "altman":  { "z_score": 4.21, "zone": "safe", "components": {} },
  "anomaly_map": { "Debtor Days": { "Mar 2015": -0.3 } },
  "red_flags": [ { "flag_type": "...", "severity": "HIGH", "first_appeared": 2019 } ],
  "peer_companies": [ { "name": "Infosys", "composite_score": 9.2 } ],
  "financial_data": { "profit_loss": {}, "balance_sheet": {} },
  "sentiment_trend": { "Mar 2020": -0.12 },
  "narrative": null
}

---

Report Dashboard

Section	Description
Fraud Score Gauge	SVG ring gauge, 0–100, color-coded by risk tier
Score Breakdown	Weighted bar chart of all 4 model contributions
Anomaly Heatmap	Industry-adjusted Z-scores for 12 ratios × 10 years — hover for exact σ
Red Flag Timeline	Chronological anomalies with severity, year, and industry context
Peer Comparison	Bar chart benchmarking against all sector peers
M-Score Analysis	Full Beneish decomposition with all 8 components
Z-Score Analysis	Full Altman decomposition with zone indicator
Financial Trend	Revenue vs Net Profit over 10 years
Sentiment Trend	VADER sentiment bar chart (positive/negative by year)
Narrative	Gemini 1.5 Flash forensic narrative — click Generate to stream
Replay Mode	Heatmap and timeline re-render year-by-year to show fraud developing

---

Satyam Case Study /satyam

An interactive reconstruction of the Satyam Computer Services fraud (2000–2009) — India's largest corporate fraud at the time, ₹5,040 Cr in fabricated cash.

• Year-by-year narrative with staggered event reveals
• Live Beneish M-Score trajectory chart showing the exact year the −1.78 threshold was crossed
• Animated metric tiles for DSRI, SGI, GMI, AQI, TATA
• Fabricated cash fill bar growing from ₹0 to ₹5,040 Cr
• Play/pause/speed controls · click any year node to jump

The model flagged probable manipulation in 2004. The confession came in January 2009.

---

Project Structure

auditgpt/
├── backend/
│   ├── api/
│   │   ├── routes.py          # All API endpoints + in-memory LRU cache
│   │   └── schemas.py         # Pydantic request/response models
│   ├── data/
│   │   ├── companies/         # Pre-cached company JSONs (TCS.json, etc.)
│   │   ├── reports/           # Pre-computed score JSONs per company
│   │   ├── auditor_notes/     # Auditor commentary data
│   │   ├── rpt/               # Related party transaction data
│   │   └── sector_summary.json
│   ├── engine/
│   │   ├── quantitative.py    # Beneish, Altman, industry Z, trend breaks
│   │   ├── scoring.py         # Composite score orchestrator
│   │   └── sentiment.py       # VADER sentiment analysis
│   ├── llm/
│   │   ├── client.py          # Gemini 1.5 Flash integration + SSE
│   │   └── knowledge_base/    # Domain knowledge for prompting
│   └── main.py                # FastAPI app entry point
├── frontend/src/
│   ├── components/
│   ├── pages/
│   ├── App.jsx
│   └── main.jsx
├── scripts/
│   ├── fetch_data.py          # Pulls raw financial data
│   └── precompute.py          # Precompute runner
└── fraud_signatures/
    └── satyam.json

---

License

MIT — built for hackathon purposes. Not financial advice. Do not use as the sole basis for investment decisions.

_{AuditGPT · NSE Forensic Intelligence · FY2026 · IAR Udaan Hackathon}