This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

SideLedger — AI-Assisted Financial Review for Freelancers

What I Built

I built SideLedger, an AI-assisted financial statement review tool designed for freelancers and independent workers.

Freelancers often receive transaction statements containing dozens or hundreds of payments, subscriptions, purchases, and repeated charges. Reviewing these manually can be time-consuming, especially when trying to understand where money went or which transactions deserve attention.

SideLedger lets a user upload a CSV transaction statement and turns it into a structured financial review.

It can:

• Categorize transactions
• Detect possible duplicate payments
• Identify recurring expenses
• Flag unusual spending
• Build a prioritized Review Inbox
• Summarize spending by category
• Answer questions about the uploaded statement through the SideLedger Assistant
• Export analyzed results

The idea is to reduce the time a freelancer spends manually inspecting transactions while keeping the final financial decisions with the user.

Rather than positioning AI as an autonomous financial adviser, SideLedger uses AI as a review assistant. Potential users could include freelancers, independent consultants, creators, and other self-employed professionals who regularly need to understand their business spending.

Demo

The demo shows SideLedger analyzing a transaction statement, displaying the financial dashboard and Review Inbox, and using the AI assistant to answer questions about the analyzed transactions.

Code

R-iddhi / hacktoberfest-expense-sorter

SideLedger - AI-assisted financial statement review for freelancers

The project is open source and the repository contains both the React frontend and FastAPI backend.

How I Built It

SideLedger uses a hybrid approach instead of asking a language model to perform every task.

Open-weight AI

I used Qwen 2.5 3B through Ollama for local language-model inference.

Qwen is used for AI-assisted transaction categorization and for the SideLedger Assistant, which answers questions using context from the uploaded financial statement.

The assistant is deliberately grounded in the analyzed statement instead of being used as a general-purpose chatbot.

Deterministic financial analysis

Python and Pandas handle operations where deterministic computation is more appropriate, including transaction processing, totals, duplicate checks, recurring-pattern analysis, and preparation of structured financial data.

This prevents simple financial calculations from depending entirely on generative AI.

Anomaly detection

SideLedger combines statistical anomaly detection with an additional tabular-model signal using TabPFN.

An IQR-based statistical method provides a simple and explainable baseline for identifying unusually large transactions, while TabPFN provides an additional signal for tabular anomaly analysis.

Human-in-the-loop review

Instead of automatically deciding that a transaction is wrong, SideLedger places suspicious or noteworthy transactions into a Review Inbox.

Users can filter the inbox by:

• Unusual spending
• Possible duplicates
• Recurring expenses

This keeps the user in control of the final decision.

Technology stack

• React + Vite
• FastAPI
• Python
• Pandas
• Qwen 2.5 3B
• Ollama
• TabPFN
• Lucide icons

The architecture separates deterministic financial processing from AI reasoning so that each component is used for the type of task it handles best.

Why Does Open Innovation Matter?

Open innovation made SideLedger possible because financial data can be highly personal.

Using an open-weight model such as Qwen through Ollama makes it possible to experiment with local inference rather than designing the entire product around sending transaction information to a proprietary hosted language-model API.

It also gives developers greater control over how the model is used, how prompts are constructed, and how AI is combined with deterministic software.

For SideLedger, open innovation is particularly valuable because the goal is not to let an AI make financial decisions. The goal is to combine transparent software, open models, statistical methods, and human review into a system where the user remains in control.

Open models also make this architecture easier for other developers to inspect, modify, reproduce, and extend.

Prize Categories

TabPFN

SideLedger integrates TabPFN as an additional tabular anomaly-detection signal alongside a statistical IQR-based method.