Industry AI systems · NoBrokerHood

Mihir Brahmaniya

GenAI Intern AI / Data Science

I build AI systems that turn messy real-world data (invoices, handwritten registers, financial statements, legacy ERP records) into structured, verified information.

Explore Projects LinkedIn

Mihir Brahmaniya pointing toward the current-role card

Interactive systems showcase

Follow the data, not the buzzwords.

Each system exposes its inputs, processing, validation and review points. Select a stage to inspect what it does.

Accounting AI

Try demo →

01Document

Ingestion & validationScanned, handwritten or digital PDFs and images enter the pipeline with type, size and page checks.

02Route

Digital vs scanned routingDigital PDFs are read from their text layer; scanned pages are rendered once to a canonical image.

03OCR ensemble

Selective OCR escalationThe primary OCR engine always runs; a second engine and an enhancement variant run only on difficult pages.

04Extraction

Contract-based LLM extractionCandidate rows are built from OCR evidence, then an LLM extracts fields under a strict per-purpose contract.

05Verification

Evidence-grounded verificationRows are verified by id against the source; a model can downgrade a row but never upgrade it.

FinSight AI

Try demo →

01Financial data

Statement ingestionTrial balance, balance sheet, P&L or income & expense statements as Excel/CSV.

02Parse

Statement detectionRows are normalised and the statement type is detected from headers and content.

03Classify

Confidence-scored classificationAccounts are classified with weighted signals: name, parent group, siblings and normal balance.

04Validate

Accounting-equation & rule engineDebit = credit, assets = liabilities + funds, and rule-based exceptions with severity.

05Risk

Materiality & ratiosMateriality-aware variance, explainable ratios, risk level and a health score.

Ledger Flash

Try demo →

01Transactions

Tenant-isolated ingestionLedger master and transactions per society, as CSV, Excel or tabular PDF.

02Memory match

Similarity retrievalEach narration is compared with reviewer-confirmed decisions; a match above 90% similarity is reused.

03Classify

Resilient batch inferenceOnly novel transactions go to the LLM, in batches, with throttling, retries and a deterministic fallback.

04Confidence

Confidence scoringEach suggestion carries a 0–100 confidence and a status: correct, possible misposting or low confidence.

05Human review

Human-in-the-loopReviewers approve, reject or correct; every action is audited.

ScrapeForge

Try demo →

01Source ERP

One browser, one sessionAn authenticated session on the legacy system; the operator logs in, the automation adopts the session.

02Automation

Browser automationNavigation, financial year and date selection through the UI, with filters read back and verified.

03Extraction

Count-validated extractionEvery page is scraped and checkpointed; the row count must match the ERP’s own total.

04Normalise

Canonical financial modelBoth systems’ reports are parsed into one canonical financial model.

05Reconcile

Explainable reconciliationPositions are matched and each difference classified: migration, system variation, presentation, roll-up or data error.

Industry projects · NoBrokerHood

Four systems, designed and implemented end-to-end.

Each project solved a real accounting or data problem. Each one gets a small interactive demo that runs in your browser on synthetic data, so you can try the core idea yourself.

01Document IntelligenceFlagship

Accounting AI

Reads scanned and digital accounting documents, extracts transactions under strict contracts, validates the arithmetic and routes uncertainty to a human reviewer.

Zero-drop candidate ledger: every source region ends in exactly one state, so nothing is silently dropped.

LangGraph
Gemini 2.5 Flash
RapidOCR · PaddleOCR
FastAPI
OpenPyXL

Explore Case Study Try Demo

System pipeline

Document
Route
OCR ensemble
Extraction
Verification
Reconciliation
Human review
Excel

02Financial Intelligence

FinSight AI

Analyses financial statements for imbalances, material movements and risk, and explains every flag with its formula, inputs and threshold.

Every flag is explainable: formula, inputs and threshold travel with each finding.

FastAPI
Pandas
Gemini 2.5 Flash
Rule engine
SheetJS · jsPDF

Explore Case Study Try Demo

System pipeline

Financial data
Parse
Classify
Validate
Risk
Insights

03Intelligent Classification

Ledger Flash

Flags transactions posted to the wrong ledger, recommends the right one with confidence and reasoning, and learns from every review.

Retrieval before generation: narrations >90% similar to a reviewed decision are resolved without an LLM call.

FastAPI
Gemini 2.5 Flash
Similarity matching
Google Sheets API

Explore Case Study Try Demo

System pipeline

Transactions
Memory match
Classify
Confidence
Human review
Learning

04Automation & Reconciliation

ScrapeForge

Automates extraction of multi-year financial reports from legacy ERPs with checkpoints and count validation, then reconciles them against the new system.

Extraction you can audit: verified filters, every page, row counts checked against the source’s own total.

Playwright · Chromium
FastAPI
Pandas · RapidFuzz
WebSockets · noVNC

Explore Case Study Try Demo

System pipeline

Source ERP
Automation
Extraction
Normalise
Reconcile
Readiness

Enterprise work: production systems, company data and proprietary details are not exposed. Demos are simplified, synthetic representations.

About · Experience

I design and build AI systems for accounting workflows where the numbers have to be right.

At NoBrokerHood I work with the Accounts team. I take each system from the problem to a working tool: AI logic, backend, frontend, data processing and automation. I build in evidence, validation and human review so that uncertainty is visible rather than hidden.

GenAI Intern

NoBrokerHood · Accounts team

23 June 2026 – Present

Bengaluru, India · On-site

Accounting AI · Document-intelligence pipeline: multi-engine OCR, contract-based Gemini extraction, LangGraph orchestration, deterministic reconciliation and human review.
FinSight AI · Financial-statement analysis: statement detection, account classification, accounting-equation and rule-based checks, materiality-aware variance and risk scoring.
Ledger Flash · Two-stage ledger classification that reuses reviewer-confirmed decisions and sends only novel transactions to Gemini in resilient batches.
ScrapeForge · Playwright-based extraction of multi-year reports from legacy ERPs with checkpointing and count validation, plus reconciliation and a readiness grade.

Designed and implemented end-to-end. AI coding and research assistants are part of my workflow; the system design, integration and debugging decisions are mine.

M.Sc. Data Science: Alliance University, Bengaluru · 2024 – 2026
BCA: Gujarat University · 2021 – 2024