LLM Project Portfolio - David Klan
EN | 日本語

Project Detail

Receipt Classification & Matching System

Automated statement-to-receipt reconciliation for faster, cleaner tax submission workflows.

  • Document AI
  • Expense Operations
  • Classification

Executive Summary

Problem

Credit card statements and receipts were reconciled manually, creating delays and inconsistent categorization before company tax submission.

Solution

Built a Gemini Colab workflow using free LLMs to match statement lines to receipts, then classify transactions with ambiguity-aware vendor handling.

Outcome

Delivered a repeatable review flow that speeds reconciliation and produces cleaner, tax-oriented expense categorization.

Use Case & Stakeholders

Finance and operations teams need reliable monthly transaction matching and categorization without scaling manual review time.

  • Finance operations analysts
  • Accounting and tax preparers
  • Business managers submitting expenses

Architecture

Statement rows and receipt text are normalized, candidate matches are scored, and ambiguous vendors are sent through LLM-assisted classification before final review.

[Credit Card Statement CSV/PDF]
           |
   OCR + normalization
           |
  Candidate match scoring
           |
Ambiguous vendor? -> LLM classifier
           |
  Human review queue + export

Tech Details

RAG / Prompt Patterns

  • Prompted ambiguity resolver with vendor-name and merchant-context hints.
  • Structured output template for category labels and confidence.
  • Lightweight fallback rules when model confidence drops.

Tools

  • Gemini Colab notebook runtime
  • Python data cleaning and matching logic
  • Free LLM endpoints for classification support

Constraints

  • Inconsistent receipt quality and OCR noise.
  • Merchant aliases and truncated statement descriptors.
  • Limited budget for model/API usage.

Tradeoffs

  • Higher recall matching can increase false positives in edge cases.
  • More aggressive categorization automation reduces manual effort but requires tighter audit checks.

Screenshots Gallery

Open full size
PDF receipt sample used for OCR extraction and line-item matching checks.
Open full size
HTML receipt representation used to validate parser and structured extraction behavior.

Lessons Learned / What I'd Improve Next

Lessons Learned

  • Confidence thresholds and review queues are essential for finance trust.
  • Merchant alias dictionaries dramatically improve initial matching quality.
  • Prompt schema enforcement prevents downstream parsing issues.

What I'd Improve Next

  • Add active-learning loops from reviewer corrections.
  • Add document-level anomaly detection for missing receipts.
  • Add monthly quality scorecards by category and vendor type.

Repro Notes / Demo Walkthrough

Non-runnable public demo: use the walkthrough steps below.

  1. Open the notebook flow and review sample statement/receipt inputs.
  2. Walk through matching score logic and ambiguity classifier prompts.
  3. Inspect exception-routing outputs and final categorized export format.