Independent product
FinMind AI: Architecting a financial agent
Designed and implemented a personal-finance MVP with a LlamaIndex-based agent that answers account, spending, and budget questions using user-specific context and tools.
Context
FinMind AI is an independent personal-finance MVP for the Bangladesh market. I created it to explore how an AI agent can work with user-specific financial information while keeping product logic, data access, and model behavior inside explicit architectural boundaries.
Problem
Financial information is fragmented across accounts, transactions, and budgets. A generic language-model response cannot answer personal questions reliably; the agent needs scoped context and controlled tools that retrieve and calculate from the user’s own data.
My role
I was the sole architect and implementer. I defined the product domain, application boundaries, data model, agent workflow, tool interfaces, retrieval approach, and provider abstractions, then implemented the MVP end to end.
Architecture decision
I chose a modular monolith with clean boundaries between the domain, application services, infrastructure, and AI module. That kept the MVP operationally simple while allowing the agent to be extracted later if its scaling or deployment needs diverge from the rest of the product.
Agent design
The LlamaIndex ReAct agent uses explicit tools for account summaries, transaction analysis, budget status, and savings questions. User-scoped context and retrieval keep the model focused on relevant financial data, while provider interfaces keep the application from depending directly on one model vendor.
Outcome
The result is a completed private MVP with a working financial agent that can answer account, spending, and budget questions through the application’s controlled tools and user-specific context.
Reflection
The central lesson was that a useful agent is an application architecture problem, not only a prompting problem. Reliability starts with constrained tools, relevant context, clear ownership of state, and testable boundaries around model behavior.