Front page

Inspiration

Over 100 million salaried individuals and first-time earners in India want to start investing. Yet, the vast majority hesitate because traditional financial platforms overwhelm them with technical jargon, aggressive trading prompts, and volatile stock recommendations without assessing whether they can afford the risk.

Most investing apps jump straight to the question: "What do you want to buy?"

We asked a more responsible question first: "Can you comfortably afford to invest right now?"

Whether it's Rahul earning ₹15,000/month needing an emergency buffer, Priya saving ₹5,000/month for higher education, or Arjun planning long-term wealth, an ordinary person needs a trustworthy companion that explains trade-offs, quantifies downside risk, and respects their financial safety net. This inspired us to build PaisaPath AI — an intelligent financial companion that guides users from their first salary to sustainable financial freedom.

What it does

PaisaPath AI is a Goal-to-Portfolio Financial Agent built specifically for first-time and retail investors:

Safety Net & Affordability Diagnosis: Evaluates monthly take-home income, living expenses, and debt obligations. It calculates an affordable monthly contribution range: $$ \text{Monthly Surplus} = \max(0, \text{Income} - \text{Expenses} - \text{Debt}) $$ Instead of recommending the entire surplus, it keeps a buffer (suggesting 40%–75%) and checks emergency fund coverage against a 6-month target: $$ \text{Target} = \text{Expenses} \times 6 $$ If high-interest debt or emergency shortfalls exist, it flags: "Safety Net First: Clear debt and build a buffer before market risk."
Deterministic Goal Planner: Users define targets (e.g., ₹2,00,000 in 3 years). It compares a zero-risk capital preservation baseline with market scenarios (4%, 7%, 10%) and computes real purchasing power using inflation compounding: $$ \text{Future Cost} = \text{Target} \times (1 + i)^n $$
Downside-Stress Portfolio Simulator: Compares Accessible (80% cash/liquid, 20% debt), Balanced, and Growth-oriented (70% equity) scenarios. Crucially, it models a severe market drawdown (down to (-35\%) equity shock) so users see that in a crash, value can drop below their principal.
AI Portfolio Health Monitor: Users enter their holdings to compute the Herfindahl-Hirschman Index (HHI): $$ \text{HHI} = \sum_{i=1}^{N} \left(\frac{w_i}{\sum w}\right)^2 $$ Categorizes diversification as Well Diversified ($\text{HHI} < 0.22$), Moderate Concentration, or Highly Concentrated ($\text{HHI} > 0.40$), automatically flagging excessive single-holding exposure (>35%).
Live AMFI Mutual Funds Feed: Integrates live published NAVs across 10,000+ Indian mutual funds with historical trend charting and honest date labeling.
Tavily News Intelligence with Causality Verification: Searches recent Indian financial news, tags publication freshness ($\le 48\text{ hrs}$ vs. older), and instructs the AI agent to explicitly verify whether news headlines causally explain market movements or remain unproven.
Multilingual Regional Coach: Explains financial concepts (SIPs, debt funds, FDs, risk tolerance) in plain, jargon-free English, हिन्दी (Hindi), and தமிழ் (Tamil).

How we built it

We designed PaisaPath AI around a strict architectural philosophy: Generative AI should never do financial arithmetic.

Reasoning AI: NVIDIA Nemotron 3 Ultra (nvidia/nemotron-3-ultra-550b-a55b via Nebius Token Factory / NVIDIA NIM). Nemotron receives verified JSON calculations and retrieved news, reasoning through liquidity, downside risks, and actionable steps.
Financial Calculation Engine: Pure Python (financial_engine.py) with deterministic mathematical functions verified by a 20-test pytest suite.
Backend: FastAPI (api.py) exposing clean REST endpoints for affordability, goal scenarios, portfolio health, AMFI lookups, and AI audits.
Market Data: Direct integration with AMFI India (NAVAll.txt) for published mutual fund NAVs and mfapi.in for historical NAV trends, complete with local disk caching and synthetic fallbacks.
News Agent: Tavily Search API (news_agent.py) with custom publication timestamp reconciliation.
Persistence: SQLite (db.py) storing user financial profiles, saved goals, and audited portfolios.
Frontend: Dual interfaces — an all-in-one Streamlit dashboard (app.py) with Plotly visualizations, plus a modern React + Vite SPA (frontend/) built with Recharts and Lucide icons.
Deployment: Multi-stage Docker container (Dockerfile, docker-compose.yml) ready for Nebius AI Cloud compute instances.

Challenges we ran into

Eliminating LLM Arithmetic Hallucinations: Early prototypes showed that language models frequently hallucinate compound interest and SIP returns. We completely decoupled arithmetic from generation: Python calculates the exact rupee values, and Nemotron only explains the trade-offs using injected JSON context.
Parsing Official AMFI Data Feeds: AMFI India publishes an unaligned text file of over 15,000 mutual fund records. Changes in plan types (Direct vs. Regular) and dividend options caused column index shifts that broke standard parsers. We engineered a resilient delimiter parser that accurately extracts the scheme code, plan, option, published NAV, and date.
The News Causality Dilemma: In financial media, it is common to falsely attribute a price dip to an unrelated headline. We built publication date reconciliation in news_agent.py to compare news timestamps against market dates, prompting Nemotron to state when a causal link cannot be established.
Resilient Offline Fallbacks: External financial APIs can be slow or encounter rate limits. We implemented disk caching and synthetic graceful fallbacks across AMFI, Tavily, and Nemotron so the UI remains interactive and never crashes during a demo.

Accomplishments that we're proud of

100% Deterministic Financial Accuracy: All savings projections, inflation compounding, and HHI calculations pass 20 automated unit tests.
Live NVIDIA Nemotron 3 Ultra Inference: Real runtime calls executing complex financial reasoning in under 3 seconds.
Honest Fintech Principles: Clearly distinguishing between periodically published NAVs (AMFI), delayed demo stocks, and verified news timestamps — never pretending delayed data is live.
End-to-End Multilingual Support: Empowering tier-2 and tier-3 Indian investors to ask financial questions in Hindi and Tamil and receive culturally relevant, beginner-friendly explanations.
Production-Ready Dockerization: A single multi-stage Docker build packaging the FastAPI backend, React frontend, SQLite persistence, and Streamlit dashboard.

What we learned

Structured Context Injection: Feeding strictly formatted JSON data into NVIDIA Nemotron produces vastly superior financial explanations compared to free-form prompting.
Indian Retail Investor Psychology: First-time investors care far more about capital preservation and liquidity ("Will I lose my money if I need it next year?") than maximizing theoretical alpha.
API Orchestration: How to seamlessly harmonize official regulatory feeds (AMFI), modern search tools (Tavily), deterministic math engines, and cloud AI inference.

What's next for PaisaPath AI

Account Aggregator (AA) Integration: Connect with RBI-regulated Account Aggregators to automatically pull bank statements and EPF balances with user consent.
Live Broker Execution Feeds: Integrate Upstox / Zerodha Kite Connect APIs for real-time stock prices and paper trading.
Voice AI Interface: Enable voice-in, voice-out interactions in regional dialects to serve semi-literate and rural households across India.
Tax Harvesting & Rebalancing Alerts: Automate notifications for long-term capital gains (LTCG) thresholds and portfolio drift rebalancing.