Carbonly.ai

The Problem I Wanted to Solve

Carbon reporting requires extracting data from electricity bills, gas invoices, and water statements across hundreds of suppliers. Each utility has a different format. Manual extraction takes hours and introduces errors that auditors catch.

What I Built

A 7-phase multi-agent system that handles the complete pipeline:

1. Classification Agent - Identifies document type (electricity, gas, water, waste)
2. Vision-to-Text Agent - Uses GPT-4o vision to read the document structure
3. Extraction Agent - Pulls out specific fields (kWh, billing period, meter number)
4. Validation Agent - Checks extracted data for logical errors
5. Normalisation Agent - Converts units and periods to standard format
6. Emission Calculation - Applies the correct emission factors
7. Audit Trail - Maintains chain of evidence for auditors

Technical Approach

Unlike template-based OCR (like AWS Textract queries), this approach uses LLM reasoning to understand documents regardless of layout changes. When a utility provider updates their invoice design, the system adapts without requiring template updates.

What I Learned

Building Carbonly taught me that document processing isn't just about extraction accuracy—it's about auditability. When a regulator asks "where did this number come from?", you need to trace it back to the source document with full confidence. That insight shapes how I approach any document automation project.