FinLearn

Inspiration

Financial reconciliation is notoriously manual, tedious, and prone to human error. CFOs and finance teams spend countless hours downloading CSVs, matching transactions, and spotting outliers in massive spreadsheets. We wanted to build more than just a chatbot—we wanted a deeply integrated, autonomous agent that acts as a true digital CFO. We were inspired by the idea of an agent that doesn't just process data but learns and adapts to an organization's specific financial footprint over time through human-in-the-loop feedback.

What it does

FinLearn is an autonomous financial context agent. You simply upload a CSV of your monthly transactions, and FinLearn's backend instantly processes it to detect statistical and semantic anomalies (e.g., your AWS bill unexpectedly jumping, or an unusual vendor category).

It presents these anomalies on a beautiful React dashboard, complete with chain-of-thought AI reasoning explaining why the transaction was flagged. A human CFO can then review the flags and correct them (e.g., "This $8,000 AWS bill is normal because we just scaled our servers"). The agent immediately absorbs this feedback into its persistent JSON memory, dynamically updating its statistical thresholds and vendor patterns. Additionally, through integration with AO Studio, the agent can spawn autonomous sub-tasks, like drafting vendor emails or generating executive summaries based on the detected anomalies.

How we built it

Frontend: A sleek, dynamic dashboard built with React (Vite) and Lucide Icons, providing deep observability into agent memory and anomalies. Backend: Node.js + Express, handling CSV ingestion, local memory persistence (memory.json), and routing. AI & Reasoning: We utilized the powerful GLM-4.7-Flash model via the TensorMux API to provide robust, chain-of-thought financial reasoning and semantic anomaly detection. Autonomy: We integrated the ao daemon and local SQLite (ao.db) to orchestrate independent, autonomous sub-agent sessions directly from the UI.

Challenges we ran into

One of our biggest hurdles was seamlessly bridging statistical anomaly detection (like Interquartile Range math) with semantic LLM reasoning. LLMs are great at text, but they can struggle with raw numerical processing over large datasets. We solved this by using the Node.js backend to perform the heavy statistical lifting (calculating typical ranges based on historical memory) and then feeding those highly specific, filtered anomalies to GLM-4.7-Flash for deep, contextual explanations. We also had to design a safe, robust way to pass React state variables into the dynamic SVG charts without causing rendering crashes during live memory updates.

Accomplishments that we're proud of

We are incredibly proud of the Self-Healing Memory System. Most LLM apps are stateless and forget everything when you refresh the page. FinLearn builds a persistent, evolving JSON knowledge graph of vendor patterns and thresholds that genuinely gets smarter the more you use it. We are also extremely proud of successfully integrating the ao CLI architecture, allowing our app to spawn entirely isolated, autonomous sub-agents that operate in their own git worktrees without breaking the main application state.

What we learned

We learned that achieving "True Autonomy" isn't about replacing the human; it’s about perfectly placing the human in the loop. By designing our system to require minimal, high-leverage feedback from a CFO, the agent can handle 99% of the tedious reconciliation while remaining safely bounded by human corrections. We also deepened our understanding of integrating local daemon orchestrators (like ao) with modern web frameworks.

What's next for FinLearn

In the future, we plan to eliminate CSV uploads entirely by integrating directly with banking APIs (like Plaid or Stripe). We also want to expand the AO integration so that sub-agents can autonomously take actions—such as actually sending the drafted emails to vendors, moving funds, or opening Zendesk tickets for accounting discrepancies, moving FinLearn from an observational tool to a fully proactive autonomous CFO.

Built With

• ao
• css3
• express.js
• react
• tensormux-api