🏦 UniBank — Banking, Simplified by AI

🌟 Inspiration

Managing personal finances can be stressful, especially when traditional banking apps are cluttered, slow, or confusing. Many people — particularly those new to digital banking — find it difficult to locate the right feature or understand how to perform simple actions like paying bills or transferring funds.

We were inspired by how natural language models like ChatGPT and Copilot make complex tasks effortless through conversation. What if banking could feel just as intuitive? That question became the foundation for UniBank — an AI-powered, human-in-the-loop assistant that makes financial management as easy as talking to your bank.

💡 What We Built

UniBank is an intelligent, conversational banking assistant that lets users manage their accounts through simple chat requests.

Instead of navigating multiple menus, users can just say things like:

“Paga el mínimo de mi tarjeta y transfiere 200 pesos a mamá.”

UniBank:

Understands the request using a planning LLM.
Breaks it down into secure, verifiable steps (e.g., verify beneficiary, pay bill, transfer funds).
Asks the user for approval before executing any action.
Executes the approved steps safely through the bank’s backend tools.

It’s like having your own personal banker — one who explains what they’re doing, waits for your approval, and never makes a mistake.

⚙️ How We Built It

Frontend: Vite + React + TypeScript — a responsive chat-like interface with approval cards and live execution updates.
Backend: FastAPI (Python) — manages sessions, validation, and tool execution with strict idempotency and audit logging.
AI Integration: Ollama running locally with Llama 3 — generates secure, structured plans in JSON and never executes without user consent.
Tool Layer: Custom adapters for safe actions like VERIFY_BENEFICIARY, PAY_BILL, TRANSFER_FUNDS, and FETCH_STATEMENT.

🧠 What We Learned

LLMs can plan, not just chat. By constraining them to strict JSON schemas and tool catalogs, they can safely generate executable workflows.
Human-in-the-loop design is essential — AI should propose, not assume.
Transparency builds trust. Showing users each planned step (with risks and rationale) makes AI-driven banking feel secure, not mysterious.
Local LLMs like Ollama can deliver strong reasoning with full data privacy — perfect for financial contexts.

🧩 Challenges We Faced

Designing safe prompts that never leak personal or financial data.
Enforcing structured output (JSON only) — LLMs love to chat, not obey schemas!
Balancing autonomy vs. control — the model needed freedom to plan but strict limits to act.
Building a real-time approval flow between chat, backend, and execution layers.
Integrating AI explainability — showing why each step was proposed and what risks it carried.