Pebbl

Teams can see cash, but they can’t act on it fast or confidently. Pebbl combines Nessie banking data, sentiment analysis, and Solana readiness to recommend confident sweep actions in real time.

Comment

Inspiration

Treasury teams often have the same blind spot: they can see balances, but they do not have a clear, intelligent system for deciding when to deploy idle cash and where to put it. We were inspired by that gap between visibility and action. During VandyHacks, we asked: What if treasury operations felt like an AI copilot experience instead of spreadsheet triage? That idea became Pebbl: a prototype for autonomous corporate treasury management that combines banking context, market sentiment, and strategy recommendations in one place.

What it does

Pebbl helps finance teams move from passive monitoring to active capital strategy by: Ingesting account and transaction context (via Nessie integration) Showing operating balance vs. sweepable surplus Monitoring crypto market sentiment from live headlines Providing AI-guided treasury recommendations in chat Tracking feedback on recommendations to improve future responses Surfacing Solana wallet and market context to prepare for execution

How we built it

We built Pebbl as a full-stack system with: Frontend: Next.js + React dashboard and marketing flow Backend: FastAPI services for chat, sentiment, treasury logic, and auth Data Stores: MongoDB for user/session/account data, Snowflake for analytics/history AI + NLP: Google Gemini + Google Cloud NLP for recommendations and sentiment scoring Market/Data APIs: NewsAPI, CoinGecko, Solana RPC, Nessie API

Core workflow:

Pull user treasury context and market signals Analyze sentiment from headlines and user intent Generate recommendation options with risk/yield context Return an explainable recommendation in chat Store interactions and feedback for iterative learning

Challenges we ran into

Multi-source integration complexity: Aligning banking, market, sentiment, and chat context into one coherent response path Data consistency: Keeping shape and freshness consistent across MongoDB, Snowflake, and third-party APIs Prompt quality and trust: Making AI responses specific, explainable, and not overly generic Hackathon speed vs. architecture quality: Shipping fast while still keeping modular services and clean API boundaries Execution realism: Designing for eventual on-chain actions while keeping a safe, demo-friendly prototype today

Accomplishments that we're proud of

Built an end-to-end prototype that actually connects data ingestion, analytics, and AI recommendation loops Delivered an AI treasury advisor that uses contextual signals instead of static canned replies Implemented sentiment-aware strategy guidance and a feedback mechanism for improvement Created a polished dashboard and onboarding flow that makes a complex system easy to demo Designed the system to be extensible toward real execution workflows

What we learned

Great fintech UX is about decision confidence, not just data display AI recommendations become much better when grounded in structured context (history, sentiment, risk profile) Eventual production systems need robust observability and safety checks around autonomous suggestions Splitting responsibilities across services early made iteration faster, not slower Hackathons reward teams that can connect product narrative, technical depth, and demo polish

What's next for Pebbl

Add policy/approval workflows before any fund movement Implement live Solana transaction execution (beyond simulation) Introduce richer risk metrics (e.g., VaR, volatility-adjusted scoring) Expand strategy connectors across more protocols and assets Add alerting for sentiment shocks and treasury threshold events Move from polling to real-time streaming updates for operations teams

Built With

Share this project:

Updates