Abstract

This project demonstrates a Proof-of-Concept (POC) for an AI-driven stock purchasing system integrated with Zerodha. The system leverages a multi-agent workflow where individual agents analyze financial transactions, generate an investment portfolio based on user savings and risk profile, select the most suitable stocks, and execute trades with optional human approval. The project highlights how AI can augment personal finance management by automating portfolio allocation and stock selection while maintaining user oversight for final investment decisions.

Introduction

Personal finance management and investment decision-making can be complex and time-consuming. Investors often need to analyze past transactions, determine available savings, allocate funds across asset classes, and choose the optimal stocks based on market conditions. This project addresses these challenges by integrating AI agents capable of autonomously performing each step of the investment process. By connecting to the Zerodha trading platform, the system demonstrates an end-to-end automated investment workflow while preserving human control through approval mechanisms.

Methodology

The project follows a multi-agent architecture with three primary agents:

1. Transaction Analyzer Agent

Reads PDF bank statements or transaction records.

Extracts monthly savings and calculates disposable income.

2. Portfolio Generator Agent

Determines an investment allocation based on the user’s age and insured status using the “100 - age” rule.

Suggests percentages for equity (stocks), bonds, emergency fund, and insurance.

3. Stock Picker/Buyer Agent

Fetches stock market data from Zerodha and filters valid equities.

Selects the best-performing stock(s) according to predefined criteria.

Awaits human approval before executing the trade and executes order.

Workflow Design

• The system maintains a central state object that carries information across agents.
• Human in the loop(HITL) is insisted for stock purchasing; the system can simulate stock purchase without execution for testing.
• The project uses FastAPI for a web-based interface, allowing workflow visualization, human approval, and displaying executed orders.

Experiments

• Tested with sample transaction PDFs representing a single month of salary and expenses.
• Simulated multiple portfolio allocations for different ages and insurance statuses.
• Verified stock selection using live Zerodha market data, filtered to valid equities.
• Conducted approval workflow: both approval and rejection scenarios were tested.
• Evaluated the UI by monitoring real-time workflow updates and executed order details in the browser popup.

Results

• Successfully extracted monthly savings from the provided transactions.
• Generated appropriate portfolio allocations aligned with user inputs.
• Correctly identified top stock candidates based on market data.
• Human in the loop(HITL) workflow functioned as expected: approved trades were executed (or simulated) and rejected trades halted execution.
• UI displayed all workflow states clearly, including executed order details, portfolio breakdown, and pending approvals.

Sample Output Snapshot:

💰 Total Savings: ₹5000
📊 Portfolio: {'Equity (Stocks)': '₹2659.09', 'Bond Securities': '₹1431.82', 'Emergency Fund': '₹454.55', 'Insurance': '₹454.55'}
📈 Suggested Stocks: [{'tradingsymbol': '364D100926-TB', 'price': 94.75, 'investment_amount': 2659.09}]
⏳ Awaiting Approval: false
✅ Approved: true
💹 Executed Order: {'tradingsymbol': '364D100926-TB', 'price': 94.75, 'quantity': 28}

Conclusion

This POC project demonstrates the feasibility of integrating AI agents with a real-world trading platform to automate personal investment workflows. The system successfully analyzes financial data, generates portfolio allocations, selects stocks, and executes trades while allowing human intervention. Future extensions could include multi-month transaction analysis, dynamic risk profiling, support for multiple brokers, and integration with predictive market models for enhanced decision-making. This project provides a foundation for intelligent, semi-autonomous personal investment systems.