Inspiration
Most money apps look backward and tell you where your money went. We wanted an AI that looks forward, at what your next 90 days will look like, and that also checks who you are about to pay, because a Bitcoin payment to the wrong wallet can never be reversed.
What it does
- FinPilot copilot: scores your financial health on 5 weighted pillars (cash flow, debt, savings, emergency fund, spending consistency), runs a Monte Carlo cash-flow forecast (2,000 seeded paths; P10/P50/P90 bands and shortfall probability, with a backtest of its own accuracy), and answers questions about your money, separating facts from your records from recommended actions, each with confidence and simulated impact. Voice input is supported.
- What-if simulator: flip subscriptions or spending off on paper and see the 12-month effect.
- Wallet Guard (Cryptiq): paste a Bitcoin address before you pay and get ALLOW / WARN / BLOCK / UNKNOWN, a 0–100 risk score from an XGBoost model on graph features, OFAC and Ransomwhere matches, an inspectable 1–3 hop evidence graph and a plain-language explanation in English or Hindi that is checked for groundedness against the retrieved evidence.
- AI councils: seven specialist agents (Strict Auditor, Cash-Flow Realist, Debt Hawk, Fraud Sentinel, Behavioural Coach, Growth Optimist, Peer Benchmarker) read a summary of your finances, never your raw transactions, vote on your risk and record their dissent; a Chain Analyst adds wallet-graph evidence.
- Statement: a printable monthly health statement with a SHA-256 audit hash, optionally anchored on Base Sepolia via EAS. Plus an Indian tax-regime comparison and an English/Hindi/Marathi UI.
How we built it
Next.js 16, React 19, TypeScript and Tailwind CSS v4; two FastAPI (Python) services (finance copilot and wallet risk); XGBoost; TigerGraph Savanna with a NetworkX fallback; Elliptic++ labels and the OFAC list; Firebase/Firestore; NVIDIA-hosted LLMs with a NuGen second opinion and an offline rule router so no data has to leave when models are unavailable.
Challenges we ran into
Keeping LLM answers honest (facts vs. recommendations, and a groundedness check on wallet explanations), making UNKNOWN never look "safe", and bridging two services securely with a route allowlist.
Accomplishments that we're proud of
A full end-to-end product that runs live, two independent AI engines in one coherent app, and explainability at every step: confidence scores, evidence graphs, dissenting agents and an auditable statement hash.
What we learned
Explainability is what makes AI financial advice trustworthy: users act on recommendations they can check.
What's next for CryptPilot
Real (non-synthetic) wallet data, Lightning Network checks, and mobile-first offline flows.
Honest status
The demo uses seeded personas and a synthetic wallet dataset, labelled as such in the app; payments run in test mode / testnet only. Built by our team in late October 2026.
Live demo: https://cryptpilot.vercel.app (press "Explore Demo App"; open https://crypt-pilot.onrender.com/api/health first to wake the free backend) Code: https://github.com/madhavzanwar/crypt-pilot Team: Madhav Zanwar, Divyani Papalkar
Built With
- bitcoin
- fastapi
- firebase
- llm
- networkx
- next.js
- nvidia
- python
- react
- tailwind-css
- tigergraph
- typescript
- xgboost
