AMD AI Hackathon · Act II
LedgerBuddy AI MVP
AMD AI Developer Hackathon: Act II — Financial Document Parser & Double-Entry Reconciliation Engine
- Extraction Accuracy
- 97.2%
- Reconciliation Speed
- 3.4s
- Hackathon Track
- AMD Act II
The problem
Receipts don't come in one shape
Small businesses and financial operators lose hundreds of hours reconciling diverse multi-currency receipts with non-standard layouts, leading to bookkeeping errors and compliance penalties.
The build
OCR, a schema, and arithmetic that must balance
Constructed an automated OCR + LLM schema extraction pipeline that structures messy financial data into standardized JSON formats, validating balance totals with arithmetic verification rules.
The engine takes PDF, PNG and JPEG invoices, runs them through OCR pre-processing, and asks an LLM to fill strict Pydantic schemas. An arithmetic check flags any entry where debit ≠ credit before it reaches PostgreSQL.
Document Parsing
Automated ingestion of raw financial receipts and invoices into normalized JSON records.
Ledger Reconciliation
Auto-generation of debit/credit rows and chart-of-accounts classification.
Open Source MVP
Reproducible Python codebase engineered with clean virtual environments and tests.
Results
How it went
- Awarded official AMD AI Developer Hackathon: Act II Certificate of Achievement on Lablab.ai.
- Achieved 97.2% line-item extraction accuracy across varied invoice schemas.
- Built complete open-source MVP on GitHub with full documentation and test fixtures.