Agentic Finance System

An AI-powered personal finance and market insight platform with a multi-agent backend, Android notification capture, retrieval-augmented answers, and adaptive financial learning.

App Images

2. Login screen	1. Google accounts	9. Asking notification access
7. Dashboard home	6. News dashboard with market sentiment	3. News dashboard with sector pulse and headlines
4. Detailed news	5. News analysis	8. Finance dashboard
10. What if simulator	11. What if simulator	12. Agent C chatbot
15. Learning page	13. Learning page	14. Learning page

What It Does

Agentic Finance System turns raw bank and UPI notifications, live market news, and curated finance knowledge into practical guidance.

• The Android app captures notification text and forwards it to the backend.
• The backend parses transactions, classifies spending, estimates financial risk, and analyzes market sentiment.
• A RAG-powered chatbot grounds answers in a curated finance knowledge base.
• An adaptive learning engine tracks concept mastery and serves personalized finance lessons.

Why This Project Exists

Most finance tools solve only one part of the problem:

• Expense trackers store transactions but do not explain spending behavior.
• Market apps show news but do not connect it to personal finances.
• Robo-advisors focus on investments and overlook day-to-day spending patterns.
• Static learning apps do not adapt to a user's actual financial profile.

This project combines those pieces into one system that is explainable, personalized, and reliable even when external AI services are unavailable.

System Overview

The report describes a three-agent architecture running on a FastAPI backend:

• Agent A: personal finance analysis from SMS and UPI notifications
• Agent B: news and market intelligence using FinBERT, NER, and trend prediction
• Agent C: grounded decision support with RAG and chat history

Each agent uses a two-tier pipeline:

• Tier 1: deterministic rule-based or ML inference
• Tier 2: LLM reasoning through Groq LLaMA 3.3 70B with fallback to Tier 1

The system also includes a Bayesian Knowledge Tracing learning engine for finance education.

System Architecture Flow Chart

flowchart TD
	U[User]
	A[Android App\nKotlin + Jetpack Compose]
	N[Notification Listener\nSMS / UPI Capture]
	API[FastAPI Backend\nPython 3.11]

	RA[Agent A\nFinance Analyzer]
	RB[Agent B\nNews + Market Intelligence]
	RC[Agent C\nDecision Synthesizer + RAG]
	RL[Adaptive Learning Engine\nBayesian Knowledge Tracing]

	DB[(SQLite / Supabase PostgreSQL)]
	KG[(RAG Knowledge Base)]
	LLM[Groq LLaMA 3.3 70B]
	EXT[External Data\nRSS + Market Feeds]

	U --> A
	A --> N
	A --> API
	N --> API

	API --> RA
	API --> RB
	API --> RC
	API --> RL

	RA --> DB
	RB --> EXT
	RB --> DB
	RC --> KG
	RC --> DB
	RL --> DB

	RA --> LLM
	RB --> LLM
	RC --> LLM

	LLM --> API
	DB --> API
	API --> A
	A --> U

Loading

Key Features

• Automatic SMS and UPI notification parsing
• Amount, merchant, and debit/credit detection from raw text
• Transaction categorization with keyword rules and ML fallback
• Personal financial risk scoring over a date range
• Live financial news sentiment analysis
• Market trend prediction from aggregated sentiment signals
• Grounded finance recommendations with retrieved context and chat history
• Adaptive learning cards, roadmap progression, and mastery tracking

Architecture Highlights

• Android client: Kotlin, Jetpack Compose, NotificationListenerService, Retrofit, Coroutines
• API backend: FastAPI, Python 3.11, Uvicorn
• Persistence: SQLite for development, Supabase PostgreSQL for production
• Authentication: Firebase Admin SDK
• NLP and ML: FinBERT, spaCy, scikit-learn, Transformers, PyTorch
• Retrieval and chat: TF-IDF RAG with Groq LLaMA 3.3 70B
• Deployment: Docker and Render support

Repository Layout

• backend/ - FastAPI app, routers, models, RAG, learning, and ML helpers
• android/ - Android client notes and integration details
• docs/ - Architecture, roadmap, problem statement, and stack notes
• presentation/ - Presentation HTML for the project demo
• scripts/ - Helper scripts for APK and backend workflows

Local Setup

From PowerShell on Windows:

cd "c:\Users\jaini\OneDrive\Desktop\SEM-6\Software Engg\agentic-finance-system"
python -m pip install -r backend/requirements.txt
uvicorn backend.main:app --reload

The backend runs at http://127.0.0.1:8000.

Main API Endpoints

• GET / - Service status
• GET /api/health - Backend health check
• POST /api/parse_message - Parse and store a transaction
• GET /api/transactions - List stored transactions
• POST /api/analyze_finance - Summarize spending and risk
• POST /api/analyze_news - Analyze market and news sentiment
• GET /api/live_market - Fetch the current market snapshot
• POST /api/synthesize - Generate grounded finance guidance
• GET /api/learning/next-card - Fetch the next learning card

Tests

Run backend tests from the repo root:

pytest backend/tests/test_api.py

Model Utilities

To regenerate the synthetic dataset and retrain the category model:

cd "c:\Users\jaini\OneDrive\Desktop\SEM-6\Software Engg\agentic-finance-system\backend"
python generate_synthetic_dataset.py
python train_category_model.py

If backend/category_model.joblib exists, the parser uses it automatically.

Deployment

• Docker support is provided through the root Dockerfile
• render.yaml is included for Render deployment
• The backend also contains the Supabase schema and Firebase auth setup needed for production deployment

Supporting Docs

• backend/README.md - Backend-focused notes
• android/README.md - Android client notes
• docs/roadmap.md - Phase-by-phase roadmap
• README_PRESENTATION.md - Presentation notes
• report_ieee_8p.tex - Main IEEE-style project report source

Scope

This repository centers on the backend, Android integration, and supporting documentation. Generated reports and presentation artifacts are kept in the workspace unless explicitly required elsewhere.