Skip to content

Pulkit7070/TrustAI

Repository files navigation

TrustAI

An AI agent swarm that restructures rejected loans into supply-based micro-credit through a Triangular Financing Loop — powered by Paytm's MCP infrastructure.

Built on Paytm's Payment MCP Server and inspired by Paytm Prism (ranked #2 globally on Spider 2.0).


The Core Idea

Traditional lending is linear and wasteful:

User → Bank → Rejected → Dead End

TrustAI introduces an Agentic Triangular Financing Loop that restructures rejected loans into supply-based credit:

         ┌─────────────┐
         │  🧑 Borrower  │ Needs supplies/inventory (not cash)
         │  (User)       │
         └──────┬───────┘
                │ ① Loan rejected by risk model
                ▼
         ┌─────────────┐
         │  🏪 Shopkeeper│ Wants more sales, inventory growth
         │  (Merchant)   │
         └──────┬───────┘
                │ ② Funds go directly to merchant (UPI Escrow)
                ▼
         ┌─────────────┐
         │  🏦 Bank/NBFC │ Gets lower risk, structured repayment
         │  (Lender)     │
         └──────┬───────┘
                │ ③ Auto-deduction from borrower's income cycle
                └───────→ Back to Borrower (supplies received)

Instead of rejecting risky borrowers, we restructure their loans through trusted local merchants. The borrower gets supplies, the shopkeeper gets a sale, and the bank gets a lower-risk, supply-backed loan with auto-repayment.


How It Works

  1. Borrower applies for credit (e.g., Rs 6,650 for business supplies or inventory)
  2. AI Agent Swarm analyzes in parallel (<200ms):
    • Analyst Agent — GNN scores the Paytm merchant graph (21 nodes) + TCN checks 12-week behavioral stability
    • Verifier Agent — Runs 5 fraud checks + verifies item prices against market rates
  3. Instead of binary approve/reject, the Orchestrator makes a 3-way decision:
    • Approved — Direct credit (low risk)
    • Structured Financing — Triangular loop via merchant (moderate risk)
    • Rejected — Only for fraud/critical risk
  4. Disburser Agent pays the shopkeeper directly via Paytm MCP (UPI Escrow) — funds never touch the borrower
  5. Auto-repayment scheduled against the borrower's next income cycle

Architecture

┌────────────────────────────────────────────────────────────────┐
│                    TrustAI Agent Swarm Engine                   │
│                                                                │
│  ┌──────────────┐  ┌──────────────┐  ┌────────────────┐       │
│  │  Analyst      │  │  Verifier    │  │  Disburser     │       │
│  │  Agent        │  │  Agent       │  │  Agent         │       │
│  │  GNN (21-node │  │  5-signal    │  │  Paytm MCP     │       │
│  │  merchant     │  │  fraud       │  │  payment       │       │
│  │  graph) + TCN │  │  detection + │  │  execution     │       │
│  │  (12-week     │  │  price       │  │  (UPI Escrow)  │       │
│  │  stability)   │  │  verification│  │                │       │
│  └───────┬───────┘  └───────┬──────┘  └───────┬────────┘       │
│          │                  │                  │                │
│  ┌───────▼──────────────────▼──────────────────▼─────────┐     │
│  │              Swarm Orchestrator                        │     │
│  │         (Prism-style self-organizing)                  │     │
│  │   Autonomous Loan Restructuring (Triangular Loop)     │     │
│  │   SHAP Explainability | WebSocket Streaming           │     │
│  └──────────────────────┬────────────────────────────────┘     │
└─────────────────────────┼──────────────────────────────────────┘
                          │
              ┌───────────▼───────────┐
              │   Paytm MCP Server    │
              │  (Model Context       │
              │   Protocol)           │
              │  UPI Escrow Payments  │
              │  Subscription Mgmt    │
              └───────────────────────┘

Execution Pipeline

PLAN → [ ANALYZE ∥ VERIFY ] → VALIDATE → DISBURSE
         (parallel)
Stage Agent What it does Latency
Plan Orchestrator Decomposes credit request into 6 sub-tasks ~1ms
Analyze Analyst GNN credit mesh (21 nodes, 6 clusters) + TCN temporal stability (12-week) ~1ms
Verify Verifier Fraud detection (5 signals) + market price verification against current rates ~1ms
Validate Orchestrator Cross-validates GNN/TCN/Fraud -> composite risk -> 3-way decision (approve / structured / reject) ~1ms
Disburse Disburser Executes UPI Escrow payment via Paytm MCP to merchant (never to borrower) ~120ms

Total pipeline: < 200ms (sub-second, matching Paytm's Groq-powered latency requirements)


Impact and Benefits

Inputs

Data Source Description
User Financial Data Loan requests, stated income, basic KYC
Merchant Graph Data 21-node Paytm transaction mesh (UPI, Soundbox, QR, POS)
Temporal Behavior Logs 12-week income, spending, and savings time-series
Fraud Signals Velocity spikes, circular transactions, KYC gaps

Direct Benefits

Benefit Detail
Sub-200ms Underwriting Instantaneous credit decisions and restructuring matching Paytm's Groq latency requirements
Proactive Risk Mitigation Shifts lending from high-risk cash transfers to secure, supply-backed merchant inventory
Ecosystem Retention Retains historically "rejected" users within the Paytm network by mediating needs through trusted shopkeepers
Explainable AI Trust SHAP-style reasoning for every approval, rejection, or structural modification

Credit Ecosystem Trajectory

Phase Outcome
Dynamic Graph Updates Merchant trust mesh updates with every new Paytm transaction
Live Escrow Tracking Real-time fund flow tracking to the merchant via MCP
Predictive Default Modeling Anticipating repayment friction before it happens
Reduced NPAs Eliminating cash misallocation drastically lowers loan default rates

Economic & Strategic Value

  • Untapped Market — Re-captures the massive "rejected loan" demographic, turning lost leads into active supply-based credit consumers within the Paytm ecosystem
  • Structural De-risking — Risk is dynamically shifted from high-risk individuals to lower-risk operational SMBs, maximizing lending ROI and safety
  • Commercial Model — B2B2C Lending Ecosystem: merchant transaction fees, increased Paytm POS/Soundbox usage, scaled SMB loan interest
  • Future Expansion — From retail supplies to universal financing across verticals (agriculture, services, gig economy) with smart-contract repayments based on live merchant ledger data

Features

# Feature Status
1 Prism-style Agent Swarm — Self-organizing agents with shared blackboard state Done
2 Triangular Financing Loop — Autonomous loan restructuring through merchant escrow Done
3 Paytm MCP Integration — AI agents disburse via Paytm's Payment MCP Server Done
4 GNN Merchant Graph — 21-node GCN on Paytm payment channels (UPI, QR, Soundbox, POS) Done
5 TCN Temporal Stability — Causal dilated convolutions on 12-week financial series Done
6 Multi-Signal Fraud Detection — 5 independent fraud checks with severity levels Done
7 SHAP Explainability — Feature contribution analysis for every credit decision Done
8 WebSocket Live Streaming — Real-time agent log streaming during execution Done
9 Hindi Voice Input — Web Speech API for Hindi commands (matching Paytm Soundbox) Done
10 Multi-Merchant Profiles — 4 demo profiles: approved, structured, rejected, fraud Done
11 Fraud Alert System — Visual red-flash alerts for critical fraud detection Done
12 Benchmark Endpoint — p50/p95/p99 latency measurement over N runs Done
13 Graph Topology API — Live 21-node merchant graph wired to D3 visualization Done
14 Hindi/English i18n — Full interface translation matching Paytm's multilingual vision Done
15 Real-time Swarm Visualizer — Pipeline stage animation with live log terminal Done

Tech Stack

Layer Technology
Agent Engine Custom Prism-style swarm (Python, asyncio, parallel execution)
Payments Paytm MCP Server (Model Context Protocol, UPI Escrow)
ML — Graph 3-layer GCN on 21-node Paytm merchant transaction graph
ML — Temporal TCN with causal dilated convolutions (12-week series)
ML — Explainability SHAP-style feature attribution (10 features)
Backend FastAPI, WebSocket, uvicorn
Frontend React 19, Vite, Tailwind CSS, Framer Motion
Visualization D3.js (graph), Recharts (charts), Three.js
Voice Web Speech API (Hindi/English)
Streaming WebSocket (real-time agent log streaming)

Paytm MCP Integration

TrustAI uses Paytm's Payment MCP Server to enable AI agents to interact with payment APIs through structured tool calls:

MCP Tool Usage in TrustAI
paytm_initiate_transaction Disburser agent pays merchants directly (UPI escrow)
paytm_transaction_status Real-time transaction tracking
paytm_create_subscription Auto-deduction schedules for loan repayment
paytm_check_balance Pre-disbursement balance verification

The Disburser agent never transfers funds to the borrower — payments go directly to the merchant via UPI escrow, preventing cash misuse.


ML Models

Merchant GNN (Graph Neural Network)

A 3-layer GCN operating on a 21-node Paytm merchant transaction graph with nodes representing:

  • Revenue channels: UPI P2M, QR Dynamic/Static, Soundbox, POS, Online PG
  • Customer segments: Regular, New, High-Value, Seasonal
  • Financial health: Settlements, Refunds, Chargebacks, Cashflow
  • Credit signals: Postpaid usage, Loan history, Credit line utilization
  • Business indicators: Inventory turnover, Supplier payments, Operating costs

TCN (Temporal Convolutional Network)

Causal dilated convolutions analyzing 12-week financial time-series (income, spending, savings) to predict behavioral stability and repayment reliability.

Fraud Detection (5-signal)

  1. Circular transaction detection (P2P symmetry)
  2. Transaction velocity spike detection
  3. New account risk scoring
  4. Loan-to-income ratio check
  5. Merchant KYC verification

API Endpoints

Method Endpoint Description
WS /swarm/ws WebSocket — real-time agent log streaming
POST /swarm/run Full swarm pipeline (analyze + verify + disburse)
POST /swarm/analyze Credit analysis with SHAP feature importance
GET /swarm/health System health + model status
GET /swarm/profiles List available merchant demo profiles
GET /swarm/profiles/{id} Full profile data for a merchant
GET /swarm/benchmark p50/p95/p99 latency benchmark
GET /graph/topology 21-node merchant graph for D3 visualization
POST /mcp/transaction Direct Paytm MCP payment
GET /mcp/status/{id} Transaction status check
GET /mcp/log All MCP tool calls in session

Quick Start

Backend

cd backend_agents
pip install -r requirements.txt
python main.py
# -> http://localhost:8000
# -> http://localhost:8000/docs (Swagger UI)

Frontend

npm install
npm run dev
# -> http://localhost:5173

Train Models (optional)

cd backend_agents
python -m models.merchant_gnn   # Train merchant GNN
python -m models.tcn             # Train TCN stability model

Why This Architecture?

Paytm's Direction TrustAI's Implementation
Prism multi-agent swarm (#2 on Spider 2.0) Self-organizing agent swarm with parallel execution
Payment MCP Server (open-source) AI agents disburse via MCP tool calls (UPI Escrow)
AI Soundbox (11 languages) Hindi voice input + Hindi/English UI toggle
Groq partnership (sub-second AI) Full pipeline executes in < 200ms
Postpaid 2.0 (credit on UPI) Triangular financing loop with supply-based credit
Merchant-first strategy 21-node transaction graph built on Paytm payment channels
Lending expansion (rural India and urban SMBs) Restructures rejected loans through local shopkeepers

Project Structure

TrustAI/
├── backend_agents/
│   ├── swarm/
│   │   ├── engine.py          # Swarm orchestrator (Prism-inspired)
│   │   └── agents.py          # Analyst, Verifier, Disburser agents
│   ├── mcp/
│   │   └── paytm_client.py    # Paytm MCP client (payment APIs)
│   ├── models/
│   │   ├── merchant_gnn.py    # 3-layer GCN on merchant graph
│   │   └── tcn.py             # Temporal convolutional network
│   ├── api.py                 # FastAPI + WebSocket endpoints
│   └── main.py                # Entry point
├── src/
│   ├── App.jsx                # Main app with routing + i18n
│   ├── lib/
│   │   └── api.js             # API base URL config
│   └── components/
│       ├── SwarmVisualizer.jsx # Real-time swarm + voice + profiles
│       ├── DecisionEngine.jsx  # AI underwriting + SHAP dashboard
│       ├── CreditMesh.jsx      # GNN graph visualization (D3)
│       ├── Navbar.jsx          # Navigation with Hindi toggle
│       ├── UserDashboard.jsx   # Borrower interface
│       ├── ShopkeeperDashboard.jsx # Merchant interface
│       └── TCNAgentVisualizer.jsx  # TCN visualization
├── ARCHITECTURE.md             # Detailed architecture & roadmap
└── package.json

TrustAI — Built for FIN-O-HACK 2026 | AI for Small Businesses Track

About

AI Agent Swarm for Merchant Credit Decisioning — Built on Paytm MCP Server & Prism Architecture | FIN-O-HACK 2026

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors