Hayal Bank Assistant

A banking assistant has to understand natural language, yet moving money leaves no room for error. LLMs are flexible but inconsistent; banking demands certainty. During my internship I built an architecture around that tension: from scratch, in 8 days, layer by layer.

A temperature-0 router, measured against a test set, sends each message to one of 5 specialist agents (Account, Transfer, General/RAG, Loan, FX), and each agent sees only its own tools. The logged-in customer's identity is injected by code at the tool layer; the LLM never touches it. I split decisions deliberately: in a workflow the code decides, in an agent the LLM does. IBAN validation, two-sided transfers in a single transaction, the loan score and annuity instalments are all deterministic Python.

Users log in with their national ID and password, then ask for balances, send money to an IBAN, check loan eligibility, query bank documents and pull live FX rates, all in plain Turkish. RAG runs semantic search with nomic-embed-text and grounds each answer in the documents; the FX service is wrapped in timeouts, error handling and a cache. Three guardrails sit on top: claimed-action verification (it catches the model saying it did something it didn't), a language-drift (CJK) detector and a format-leak filter. Small talk takes a fast path and never reaches the LLM.