Mihir BrahmaniyaGenAI Intern · AI / Data Science
LinkedIn
All projects
Document IntelligenceIndustry Project · NoBrokerHood

Accounting AI

Document intelligence that turns invoices, handwritten registers and bank statements into verified, ERP-ready accounting data.

My roleDesigned and implemented the system end-to-end, including the AI logic, backend, frontend, data processing and automation workflows.

Explore Try Demo

Problem

Housing-society accounting runs on messy documents: handwritten petty-cash registers, scanned ledgers, bank statements, vendor bills. Keying them into the ERP is slow and error-prone, and naïve OCR-plus-LLM pipelines silently drop rows or invent numbers, which accounting can’t tolerate.

System pipeline

  1. Document
  2. Route
  3. OCR ensemble
  4. Extraction
  5. Verification
  6. Reconciliation
  7. Human review
  8. Excel

Approach

Validation is part of the system, not an afterthought.

Reads scanned and digital accounting documents, extracts transactions under strict contracts, validates the arithmetic and routes uncertainty to a human reviewer. The workflow keeps inputs, decisions and exceptions visible so that a reviewer can understand how the output was reached.

My contribution

  • Document ingestion, page rendering and OCR orchestration
  • Structured extraction contracts and evidence verification
  • Deterministic accounting checks, human review and Excel output
  • Unit, regression and golden-document benchmark suites

Try the demo

Accounting AI: interactive demo

Invoice → accounting-ready record: real text extraction, field detection, line items, GST and arithmetic validation.

  • Computed
  • Simulated visual
  • Local model
Document

Select a safe synthetic document.

Northstar FacilitiesSYNTHETIC INVOICE
Invoice
NSF-2408-118
Date
2026-08-18
GSTIN
29ABCDE1234F1Z5
ItemQtyRateAmount
Common area cleaning1₹18,500₹18,500
Cleaning supplies4₹825₹3,300

Written total ₹25,724

◇

Choose a document and process it. The arithmetic checks are calculated in your browser.

This interactive demonstration is a simplified, synthetic representation inspired by an enterprise project. Production systems, company data and proprietary implementation details are not publicly exposed.

Result

A result that can be inspected.

The portfolio demo produces a structured invoice record, calculates every line, checks tax and total consistency, and routes missing or conflicting fields to review.

Technical architecture

Stage-by-stage system design

  1. 01
    Document · Ingestion & validation

    PDF / JPG / PNG → Validated file + SHA-256 manifest

  2. 02
    Route · Digital vs scanned routing

    Validated file → Text layer or canonical page image

  3. 03
    OCR ensemble · Selective OCR escalation

    Canonical page image → Lines, boxes, confidences (all kept as evidence)

  4. 04
    Extraction · Contract-based LLM extraction

    OCR evidence + candidate ledger → Structured rows with row ids

  5. 05
    Verification · Evidence-grounded verification

    Structured rows + page image → Verified / needs-review rows

  6. 06
    Reconciliation · Deterministic accounting

    Verified rows → Reconciliation checks

  7. 07
    Human review · Human-in-the-loop

    Review items → Approved record

  8. 08
    Excel · Formula-safe export

    Approved record → ERP-ready XLSX

Tech stack

Tools used in the production system

  • LangGraph
  • Gemini 2.5 Flash
  • RapidOCR · PaddleOCR
  • FastAPI
  • OpenPyXL

The portfolio demo itself is static TypeScript running locally in the browser with synthetic data and no API key.