An AI-powered Audit Readiness Assistant combining deterministic analytics, explainable AI, and human judgement.
During the past few years I have spent a great deal of time studying explainable AI, data science, and decision-support systems. One question kept coming back:
Could AI make financial review less intimidating without pretending to replace professional judgement?
BeLedgerReady is my exploration of that idea.
Rather than producing opaque risk scores, the application builds deterministic evidence first, then uses AI to explain those results in clear language. The objective is not to automate an audit---it is to help people understand what deserves attention and why.
- Upload and validate ledger CSV files
- Guided column mapping
- Dataset profiling
- Deterministic statistical, ledger, and control analyses
- Evidence-based Findings Explorer
- Optional AI explanations grounded in deterministic evidence
- Executive dashboard
- Professional printable report
Upload CSV
↓
Map Columns
↓
Dataset Profile
↓
Deterministic Analysis
↓
Evidence
↓
Findings
↓
(Optional) AI Explanations
↓
Professional Report
AI sits deliberately at the end of the pipeline.
- Deterministic detectors generate Findings and Evidence.
- Only those structured outputs are sent to OpenAI.
- AI explains the evidence---it does not create it.
The model never analyses raw ledgers directly and cannot invent findings, modify severity, or change confidence scores.
Note: AI explanations are generated on demand. If you want them included in the exported report, generate the explanations before exporting.
cd backend
py -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e .
.\.venv\Scripts\python.exe -m uvicorn beledgerready.main:app --reload --app-dir srccd frontend
npm install
npm run devFrontend: http://localhost:3000
Backend: http://localhost:8000
Create:
backend/.env.local
Example:
OPENAI_API_KEY=your_openai_api_key_here
OPENAI_MODEL=gpt-4.1-miniThis file is intentionally ignored by Git and must never be committed.
Want to explore BeLedgerReady in just a few minutes?
This repository includes two fully synthetic demonstration datasets in the sample-data/ folder.
| Dataset | Best for |
|---|---|
synthetic-ledger-small.csv |
A quick tour of the application and workflow |
synthetic-ledger-105-rows.csv |
A richer demonstration with more findings and evidence |
Both datasets are entirely synthetic and contain no real financial or personal information.
This project was developed during OpenAI Build Week using an iterative human--AI engineering workflow.
I remained responsible for the product vision, architecture, detector design, user experience, validation strategy, testing, and final decisions.
OpenAI Codex assisted with:
- implementing structured engineering work orders;
- frontend and backend implementation;
- refactoring;
- debugging;
- automated test generation;
- UI polishing;
- documentation improvements;
- validation support.
GPT‑5.6 acted as an engineering partner throughout the project by helping break complex work into manageable phases, reviewing architectural choices, refining prompts, and improving documentation and usability.
Every significant design decision remained under human supervision.
- Deterministic evidence before AI
- Explainability before automation
- Human review remains central
- Traceability over black-box scoring
- Local-first processing
- An accounting package
- An ERP
- An audit opinion
- A fraud detector
- A replacement for professional judgement
Built for the OpenAI Build Week Hackathon.
Special thanks to the OpenAI team for making Codex and GPT‑5.6 available during the event and enabling an iterative, engineering-first development workflow.