-
-
ReconcileAI ingests bank statement and accounting records — six transactions each, ready to compare.
-
The engine catches an amount mismatch (Amazon) and two missing transactions in one pass — no manual spreadsheet diffing.
-
After human-approved corrections: mismatches go from 1→0, missing items from 2→1 — with a clear action log and AI commentary.
-
The agent's actual toolset — Strands Agents SDK tools for reconciling, investigating, prioritizing, and proposing fixes.
-
The core reconciliation logic is plain deterministic Python — numbers are never left to the AI to compute.
-
Python owns facts, AI owns reasoning': the AI only summarizes unresolved issues the code has already verified
-
Nothing gets written back to the books without an explicit human-approved proposal passing through here first.
-
Runs on a local qwen2.5:7b model via Ollama, no cloud LLM dependency required for the reasoning layer.
Inspiration
Financial reconciliation can be repetitive and time-consuming. I wanted to build something that could help people find and understand financial discrepancies faster.
I was also interested in learning how AI agents can be used to solve a real-world problem, not just answer questions.
What it does
ReconcileAI compares bank transactions with accounting records and finds discrepancies.
It can identify:
- Matching transactions
- Amount differences
- Missing transactions
- Transactions that need attention
The AI agent can investigate the discrepancies, prioritize them, and suggest corrections. A human must approve a correction before anything is changed.
How we built it
I built ReconcileAI using Python, Streamlit, the Amazon Strands Agents SDK, and an LLM.
I used Python for the actual transaction comparison and an AI agent for investigation, prioritization, and generating correction proposals.
Challenges we ran into
One of the biggest challenges was making the AI agent work reliably and deciding which tasks should be handled by AI and which should be handled by normal code.
I also had to deal with debugging, connecting the different parts of the application, and getting the project ready for deployment.
Accomplishments that we're proud of
I'm proud that I built a working AI agent that goes beyond a simple chatbot.
It can actually work with financial data, investigate problems, and create useful correction proposals while keeping a human in control.
What we learned
I learned that building an AI agent is not just about using an LLM.
The agent needs the right tools, clear instructions, and limits. I also learned a lot about the Strands Agents SDK, tool calling, Streamlit, Python, and building an end-to-end AI application.
What's next for ReconcileAI
The next step would be connecting ReconcileAI to real banking and accounting systems instead of CSV files.
I would also like to let it use invoices and receipts when investigating transactions and improve its ability to explain why a transaction is suspicious or needs correction.
The goal is to make reconciliation faster and easier while keeping humans in control.
Log in or sign up for Devpost to join the conversation.