ProofLedger AI — Evidence-First Financial Reconciliation
The problem
Small businesses often spend hours comparing bank statements with accounting ledgers, identifying missing transactions, checking duplicate payments, and investigating amount mismatches. This repetitive work can make month-end closing slower and harder to verify.
Financial reconciliation requires more than a plausible explanation. The calculations must be consistent, discrepancies must be traceable to source records, and decisions must remain reviewable.
Our solution
ProofLedger AI is a financial reconciliation workspace designed to make discrepancies easier to identify, investigate, and document.
Its core principle is:
Deterministic financial calculations. Evidence-based investigation. Human-controlled resolution.
The current implementation focuses on reliable reconciliation and transparent, rule-based analysis rather than depending on an external AI model.
How it works
1. Import and reconcile transactions
Users can import bank statement and accounting ledger CSV files. The application parses the records in the browser and uses a deterministic TypeScript engine to match transactions, identify unmatched entries, detect potential duplicates, evaluate date differences, and calculate amount discrepancies.
Financial amounts use integer minor units, such as paise for INR, to avoid floating-point rounding errors.
2. Investigate discrepancies
The workspace provides an overview of reconciliation results, financial variances, transaction matches, and outstanding exceptions. Users can inspect the relevant bank and ledger records to understand why a transaction requires further review.
For example, the fictional Northstar Studio demonstration contains a ₹49,760 bank credit matched against a ₹50,000 ledger entry, producing an amount difference of ₹240.
3. Keep analysis transparent
The current demo uses deterministic calculations and local rule-based classifications to provide initial investigative guidance. It does not currently integrate a working NVIDIA model or provide live AI-generated analysis.
Suggestions are intended to guide further investigation, not establish that a suspected fee, duplicate, or missing transaction has been conclusively explained.
4. Preserve human control
Users can approve, reject, or defer an exception, add review notes, and export reconciliation information in CSV, JSON, and Markdown formats. These actions support review and documentation rather than automatically changing accounting records or transferring money.
5. Protect sensitive financial information
The application parses imported CSV records in the browser and includes privacy-oriented safeguards, including a guest workflow, sensitive-identifier redaction in the investigation payload, and protection against spreadsheet formula injection in exports.
The investigation workflow includes a consent step before submitting its minimized payload to the project server. This does not mean every possible privacy risk has been eliminated; users should review the information being submitted before granting consent.
The technology
The project uses React, TypeScript, Vite, Node.js, Express, PapaParse, and Zod. It includes automated tests covering important areas such as financial arithmetic, reconciliation logic, CSV imports and exports, and privacy-related behavior.
Challenges and what I learned
A central challenge was separating financial calculations from investigative interpretation. Transaction dates can use different formats, debit and credit conventions must remain consistent, and possible matches must be evaluated without silently treating uncertain results as confirmed facts.
This project reinforced the value of deterministic calculations, traceable evidence, defensive input validation, and explicit human review when developing software for financial workflows.
Development transparency
ProofLedger AI is an existing codebase being presented for this competition. The capabilities described above refer to the current implementation; they are not automatically claimed as features newly developed for LaunchPad Hacks.
Any additional work completed specifically for the competition will be documented separately, with verifiable changes and test evidence. Future AI integration is not presented as an implemented capability.
Current limitations
The application currently supports CSV-based reconciliation; PDF and scanned-statement extraction are not supported. The investigative guidance in the current demo is rule-based, and the application does not use a live external AI model.
The demonstration uses synthetic financial records and is intended as a reconciliation aid, not a replacement for professional accounting judgment.
ProofLedger AI aims to make financial reconciliation more transparent, reviewable, and privacy-conscious by putting calculations and evidence first.
Built With
- artificial-intelligence
- csv
- cybersecurity
- data-visualization
- express.js
- node.js
- nvidia-nemotron
- papaparse
- react
- supabase
- typescript
- vite
- vitest
- zod
