BARAI

AI-powered carbon activity intelligence from business documents.

Barai uses Claude to interpret invoices, bills, receipts and other business documents and convert their contents into structured carbon activity data.

Explore architecture Claude integration

Documents

Extraction

Claude

Activity Data

Emission Factors

Emissions

Problem

Carbon accounting depends on activity data such as:

Fuel consumption
Electricity consumption
Purchased materials
Transportation
Logistics activity
Business travel
Waste
Other operational activities

However, the underlying information frequently exists inside:

Invoices
Bills
Receipts
Purchase orders
Logistics documents
Vendor documents
PDFs
Scanned documents

These documents contain useful information but are not structured for direct carbon accounting.

Manual interpretation creates:

High operational effort
Inconsistent categorization
Unit inconsistencies
Difficulty identifying relevant activities
Difficulty scaling carbon data collection

Barai addresses this document-to-activity-data gap.

Solution

The core concept:

Unstructured Business Documents
              ↓
       Raw Extraction
              ↓
    Semantic Interpretation
              ↓
    Structured Activity Data
              ↓
   Emission Factor Mapping
              ↓
   Carbon Calculation

Barai separates **unstructured understanding** from **deterministic carbon calculation**.

This is an important architectural principle. The system uses AI where semantic reasoning adds value, and deterministic software where calculation precision is required.

Unstructured Documents

Raw Extraction

Semantic Interpretation

Structured Activity Data

Emission Factor Mapping

Carbon Calculation

How It Works

A detailed pipeline for transforming business documents into carbon emission data.

Step 1 — Document Upload

A user uploads an invoice, bill, receipt, or other relevant business document.

Step 2 — Document Extraction

OCR/document processing extracts the raw information.

Item: Diesel Fuel
Quantity: 450
Unit: L
Amount: ₹42,300
Vendor: ABC Fuels
Date: 12 August 2026

Step 3 — Claude Interpretation

Claude receives the relevant extracted information and determines what the information represents.

It may need to resolve:

What the item actually represents
Whether it is relevant to carbon accounting
Activity category
Quantity
Unit
Context
Ambiguity
Missing information
Confidence

Step 4 — Structured Activity

Produce normalized activity data.

{
  "activity": "diesel_consumption",
  "quantity": 450,
  "unit": "litre",
  "category": "stationary_or_mobile_fuel",
  "source": "invoice",
  "confidence": 0.96
}

Step 5 — Validation

Validate:

Required fields
Data types
Units
Quantity ranges
Activity schema
Category validity

Invalid or ambiguous results should be flagged for additional processing or human review.

Step 6 — Emission Factor Matching

Match the normalized activity against an emission-factor dataset.

Step 7 — Calculation

Perform deterministic emission calculations. The LLM should not be responsible for arithmetic that can be performed deterministically.

Claude as the Interpretation Layer

Barai uses Claude to bridge the gap between raw document extraction and structured carbon activity data.

Traditional OCR can identify text, numbers, and fields. It does not reliably determine the semantic meaning of those fields in the context of carbon accounting.

Claude is used to interpret the extracted information and map it into a controlled activity schema.

📥Business Document

⚙️OCR / Parsing Layer

📊Raw Extracted Data

🧠Claude

📊Structured Activity

⚙️Validation

📊Emission Factor Data

⚙️Deterministic Calculation

Why Claude excels here

Semantic interpretation

Understanding what a line item means in the context of carbon accounting.

Context resolution

Using surrounding document information to interpret ambiguous fields.

Normalization

Converting inconsistent descriptions into a controlled activity representation.

Classification

Mapping activities to an appropriate carbon-accounting category.

Confidence estimation

Identifying uncertain interpretations that may require validation or human review.

Example

A concrete before/after example showing the transformation from document to structured activity data.

Input document

invoice.txt

ABC FUELS
Invoice #INV-10482

Diesel Fuel       450 L       ₹42,300
Transportation    1 service   ₹8,000

Date: 12 Aug 2026

Raw extraction

extraction.json

{
  "items": [
    {
      "description": "Diesel Fuel",
      "quantity": 450,
      "unit": "L",
      "amount": 42300
    },
    {
      "description": "Transportation",
      "quantity": 1,
      "unit": "service",
      "amount": 8000
    }
  ]
}

Claude interpretation

interpretation.json

{
  "activities": [
    {
      "activity": "diesel_consumption",
      "quantity": 450,
      "unit": "litre",
      "confidence": 0.96
    },
    {
      "activity": "transportation_service",
      "quantity": 1,
      "unit": "service",
      "confidence": 0.72
    }
  ]
}

Note: The second activity may require additional context before it can be reliably mapped to an emission factor. This demonstrates why semantic interpretation matters.

Technical Architecture

A clean architecture showing the full system flow from document upload to carbon report.

👤User

⚙️Document Upload

⚙️Document Processing

⚙️Raw Extraction

⚙️Interpretation Layer

🧠Claude

⚙️Validation

📊Activity Data Store

⚙️Emission Factor Engine

⚙️Emission Calculation

📄Carbon Report

Design Principles

Engineering principles that guide Barai's architecture.

LLM for interpretation, not arithmetic

Use Claude where semantic reasoning is valuable. Use deterministic software where deterministic logic is sufficient.

Structured outputs

Claude responses should conform to a predefined schema.

Validation

Never blindly trust model output.

Confidence-aware processing

Low-confidence interpretations should be flagged rather than silently accepted.

Traceability

Activity data should retain its relationship to the original source document and extracted fields.

Human review

Ambiguous or low-confidence activities should be reviewable.

Current Status

Current stage

Barai is currently being developed as a document-to-carbon-activity intelligence system.

Current technical focus:

Document extraction
Activity interpretation
Claude integration
Structured activity schemas
Validation
Emission factor matching
Carbon calculation pipeline

Do NOT claim:

Production customers
Revenue
Production-scale usage
Proven accuracy
Existing enterprise deployments

unless those claims are actually provided later.

Roadmap

Short-term development roadmap.

01Document ingestion

02Reliable extraction

03Claude-powered activity interpretation

04Validation and confidence scoring

05Emission factor matching

06Carbon calculation

07Human review workflows

08Enterprise reporting

Problem#

Solution#

How It Works#

Claude as the Interpretation Layer#