Month Proof
Month Proof automates month-end close for US finance teams.
Core trick
Computes all variances in pandas and blocks saving the LLM-written narrative unless its figures match within 2% tolerance.
Tags
Description
Month Proof automates month-end close for US finance teams. Upload your financial Excel files → the agent reads them, compares to history, flags anomalies, writes a plain-language report, and sends it via email — all in under 2 minutes.
Finance teams spend 10-15 hours/month on manual variance analysis and narrative writing. Month Proof reduces that to zero human-hours for the first draft.
The system uses a 3-agent pipeline: Parser (column mapping with Haiku), Comparison (pure pandas variance calculation), and Interpreter (narrative generation with Opus 4.7). The critical innovation is the numeric guardrail: Claude never does math — all calculations happen in Python, and a verification layer ensures the narrative matches the pandas output within 2% tolerance. No report is saved until the guardrail passes.
Built with FastAPI, React, Supabase, and Claude Opus 4.7. Tested with real GL exports from NetSuite, QuickBooks, and manual consolidation scenarios. The entire codebase was developed using Claude Code as the primary development engine.
README excerpt
# TrueCost ### AI-Powered Month-End Close Agent *Built with Claude Opus 4.7 — Anthropic Hackathon April 2026* > Now documented at [jansoganci.com/work/month-proof](https://jansoganci.com/work/month-proof) — part of my public build log. --- ## What it does Drop your messy financial Excel files. TrueCost reads them, compares them with history, finds anomalies, writes a plain-language report, verifies its numbers, and exports a close package. Built for US finance teams spending 10-15 hours/month on manual close work. --- ## How it was built with Claude Code This project was developed entirely using Claude Code as the primary development engine. Claude Code was used to: - Architect the agent orchestration loop (tool-use pattern) - Write and debug all pandas parsing logic - Design the numeric guardrail system - Generate and iterate on narrative prompts - Fix edge cases in NetSuite XML format detection - Write FastAPI endpoints and Supabase integration To see Claude Code in action on this codebase, run: ```bash claude "explain how the numeric guardrail works in backend/tools/guardrail.py" claude "add support for .xlsm files in backend/tools/file_reader.py" claude "write a test for the variance calculation in backend/agents/comparison.py" ``` --- ## Quick Start ### Prerequisites - Python 3.11+ - Node.js 18+ - Supabase account - Anthropic API key ### 1. Clone and install ```bash git clone https://github.com/jansoganci/iron-ledger.git cd iron-ledger cp .env.example .env