Inspiration
As UCSC students, we know the feeling: part-time job income, student loans, credit card debt, and no clear plan for how to make it all work. Financial planning tools either talk down to you or assume you already know what you're doing. We wanted something that actually meets students where they are — no jargon, no judgment, just a real plan.
What It Does
Pathwise is an autonomous AI agent that turns your financial situation into a personalized month-by-month roadmap. You describe your finances in plain English — income, debts, expenses, goals — and Pathwise does the rest.
It autonomously:
- Fetches live financial rates (HYSA, mortgage, S&P 500) in real time using Tavily
- Extracts your financial profile using NVIDIA Nemotron via OpenClaw
- Reasons through your priorities in the right order: emergency fund → high-interest debt → investing → goals
- Generates a 12-month roadmap with specific dollar amounts for each phase
- Saves your plan to persistent memory across sessions
- Runs what-if scenarios on demand ("what if I get a raise?" or "what if I pay $200 extra toward debt?")
The web UI includes interactive charts, a paycheck breakdown tool, financial milestone tracking, an investing basics guide, and a future net worth projector.
How We Built It
Agent Framework: OpenClaw 2026.5.12, configured with NVIDIA Nemotron as the reasoning engine. Every AI call in the app routes through the OpenClaw agent, which handles multi-step reasoning and persistent memory across sessions.
LLM: NVIDIA Nemotron (llama-3.3-nemotron-super-49b-v1) accessed via NVIDIA NIM endpoints at build.nvidia.com. Nemotron's function-calling capabilities made it ideal for returning structured JSON financial plans.
Live Data: Tavily Search API for real-time rate lookups — HYSA rates, 30-year mortgage rates, and S&P 500 outlook — so every plan is grounded in current market conditions.
Web UI: Flask backend serving a single-page app with Chart.js for interactive visualizations including stacked bar charts for surplus allocation, line charts for portfolio growth, and donut charts for paycheck breakdown.
Memory: JSON-based persistent storage that saves your financial profile and plan across sessions, enabling the what-if engine to reference your existing roadmap.
Challenges We Faced
The biggest challenge was integrating OpenClaw into a web application. OpenClaw is designed as a conversational agent framework, not a programmatic API — routing structured JSON requests through it required careful prompt engineering and subprocess management. We also had to work around the event WiFi blocking clawhub.ai, which prevented NemoClaw installation and forced us to use the embedded agent fallback.
Prompt engineering for reliable structured JSON output was another major challenge. NVIDIA Nemotron is powerful but needed precise instructions to consistently return valid JSON without markdown fences or extra text. We built a robust parser with retry logic and graceful error handling.
Finally, merging two teammates' parallel UI and agent development required careful conflict resolution under time pressure.
What We Learned
- How to build a genuinely autonomous agent using OpenClaw and NVIDIA Nemotron
- How to integrate live tool use (Tavily) into an agent workflow for real-time data
- The importance of structured output design when building AI-powered applications
- How persistent memory transforms a one-shot chatbot into a true agent that knows your situation
- That financial planning is fundamentally a multi-step reasoning problem — exactly what Nemotron excels at
What's Next
- Multi-session memory that tracks progress month-over-month
- UCSC-specific financial resources (work-study optimization, financial aid gap planning)
- Mobile app so students can check their plan anywhere
- Integration with bank APIs for automatic transaction categorization
- NemoClaw deployment for privacy-hardened financial data handling
Log in or sign up for Devpost to join the conversation.