EmeilSG-Evidence-first sustainability reporting
An evidence-first ESG platform automating SEBI BRSR compliance for MEIL turning raw invoices into audit-ready reports with 100% data lineage.
- Video of it working
- Problem, approach and result explained
- Public source code
- Built during the challenge
- Answered the full brief· 13 of 13
- Team confirmed roles· 6 of 6
- Expert reviewed· 6 reviews
100%
Evidence traceability
146
SEBI questions
< 10s
Audit pack generation
Walkthrough
Overview
The problem
If you look at large infrastructure conglomerates like MEIL (Megha Engineering & Infrastructures Ltd), they operate hundreds of massive projects across the country everything from manufacturing electric buses and drilling rigs to building irrigation canals and highways. Under SEBI regulations, India’s top 1,000 listed companies are mandated to file detailed Business Responsibility and Sustainability Reports (BRSR), and now SEBI requires third-party "reasonable assurance" on BRSR Core KPIs. Here is the dirty secret of enterprise ESG reporting today: it is spreadsheet chaos. Every month, sustainability teams try to collect utility bills, diesel receipts, and contractor records from dozens of remote project managers over WhatsApp and email. Someone types a fuel number into an Excel sheet, accidentally adds a zero, uses the wrong unit of measurement, or uses an outdated carbon conversion factor. By the time Big 4 auditors arrive at year-end to verify the numbers, nobody can prove which physical bill produced the figure on row 240 of a consolidated spreadsheet. A single unsubstantiated number can trigger regulatory scrutiny, financial penalties, or allegations of greenwashing. What did we set out to do? We set out to eliminate "spreadsheet compliance" by building EmeilSG—an evidence-first BRSR reporting platform designed specifically for MEIL’s real- world structure. Our goal was simple: make reporting start with the proof, not the spreadsheet. Instead of forcing site engineers to fill out intimidating carbon calculation forms, they simply upload their real bills and invoices. The system reads them with OCR, validates the data, runs transparent and deterministic carbon calculations, and rolls the figures up into SEBI-prescribed BRSR reports where every single number can be clicked to reveal its original source document.
The approach
We built EmeilSG, an evidence-first BRSR portal. A person uploads a bill as a PDF or a phone photo. Our own document reader extracts the values and scores its confidence in each one. The uploader confirms them against the document, side by side. The portal then validates the record, calculates emissions with a versioned conversion factor, and sends it to a different person for approval. Approved figures roll up the organisation tree into the BRSR report.
Under it is a Next.js portal, a Postgres database with row-level security, and a Python OCR worker running in Docker. Access is by role and by place in the organisation: a project manager sees their project, an administrator sees their unit and everything below, nobody sees a sibling unit. Forms would have been faster to build and easier to demo. But a typed number has no evidence behind it, and that is exactly the weakness the brief is about. So upload is the only way in, and a hand-typed value is allowed only as a flagged exception with a written reason.
We built an enterprise-ready portal backed by an automated OCR ingestion pipeline, a deterministic calculation engine, and a database-locked audit trail: 1. A Living Organization Tree: We modeled MEIL’s multi-subsidiary structure (including Olectra Greentech and EVEY) across four active operating projects (like the BEST Mumbai e-buses and the Seetharampur manufacturing plant) with 9 distinct user roles. 2. Human-in-the-Loop OCR: When a site manager uploads an electricity bill or fuel slip, our background OCR engine extracts key metadata (invoice numbers, dates, consumption amounts). But we follow a strict rule: AI assists, a human decides. The user reviews the extracted values on a side-by-side verification screen before anything is committed. 3. Automated Multi-Layer Validation Engine: Before a record is saved, our engine checks for missing fields, unit mismatches, overlapping billing cycles on the same meter, duplicate invoices, and statistical anomalies (flagging consumption that jumps more than 50% above a site's 12-month rolling average). 4. Deterministic Calculation & Full SEBI Mapping: We mapped all 146 SEBI BRSR questions across Sections A, B, and Principles 1–9. Using versioned emission factors from the Central Electricity Authority and IPCC, our calculation engine records the exact formula and factor version used for every single metric. What we chose not to do: No AI-generated figures. OCR proposes, a person confirms. The chatbot only words answers; the portal computes every number with the same rules as the reports. It can also navigate the new user, No guessed values. A field the reader cannot find stays empty for a person to fill. No built-in emission factors. Each factor is entered with its source and approved by a second person, because an unsourced factor cannot be assured. No self-approval. The person who uploads is never the person who approves.
The result
It works end to end, on a live deployment, with realistic data. Coverage: all 146 BRSR questions across Sections A, B and C, with 47 reportable activities across Scope 1, 2 and 3, workforce, water, waste, safety and CSR. Demonstration: 18 months of documents (April 2025 to September 2026), about 66 bills a month, filed by five departments across four projects, giving roughly 1,000 records. Validation caught real errors: a faint headcount figure was read as 439 instead of 139; the engine failed it and HR corrected it before approval. Traceability: from any report value to the original document in three clicks. Speed: the reader takes about ten seconds a page; a year’s roll-up for 300 projects and 86,400 records returned in 0.23 seconds in our scale test. Outputs: four report templates, including the full SEBI BRSR and a GHG inventory in Excel. How we measured: we had the build independently audited twice during development. The first audit scored 52 out of 100; after closing its findings the second scored 72. We then ran security, database, accessibility and dependency audits: Measure Before After Readiness audit score 52/100 72/100 Security advisories in shipped code 14 2 (no fix available, recorded) Accessibility checks failing 36 0 Automated database tests - 482 passing End-to-end scripts - 19 of 20 passing in full
• Zero Disconnected Figures (100% Traceability): Every reported figure in our Scope 1, Scope 2, and BRSR dashboards is clickable. An auditor can click on an emission total and instantly view the original PDF invoice, the OCR confidence score, the human reviewer who approved it, and the exact math formula. • Instant Anomaly & Error Catching: During our test runs, the validation engine caught 100% of simulated billing overlaps and blocked duplicate vendor invoices that would have previously slipped through unnoticed in manual spreadsheets. • Audit Preparation from Weeks to Seconds: Instead of taking weeks to hunt down paper receipts for external auditors, our Auditor Evidence Bundle compiles all original source PDFs, spreadsheets, and cryptographic SHA-256 checksums into an immutable ZIP package with a single click. • Instant SEBI Compliance: Generates official, ready-to-file SEBI BRSR Full and Core reports in PDF format and GHG calculation workbooks in Excel.
Reflection
What we would do differently: test on real documents from day one. Most of our OCR rules were built on generated samples, and the one real electricity bill we tried exposed three layout problems in an afternoon. We would also put the OCR worker on an always-on host earlier; it currently runs in Docker on a laptop, so uploads wait when that machine is off.Passing tests can hide a broken feature. Our evidence bundle had seven green unit tests and had never worked against the real database. Only running the whole thing showed it. After that, we stopped calling anything done until it had run end to end. Refusing to guess is a feature. An empty field a person fills is worth more to an assurer than a confident wrong number. Constraints help. Taking away the data-entry form forced every other part of the design to be about evidence.
Submitted to The BRSR Reporting Portal
- Your idea in one line
EmeilSG turns scattered site invoices into audit-ready BRSR reports, with every figure traceable to its source document, built for the ESG teams of large listed infrastructure groups and the auditors who verify them.
- What will be working by the finale?
(1) OCR document pipeline with confidence scoring, validation and maker-checker approval; (2) seven-level hierarchy, shared projects, roll-up and drill-down, bulk import, 8 custom roles; (3) Scope 1/2/3 engine with versioned conversion factors; (4) BRSR Sections A to C auto-filled from approved records, indicator mapping, readiness view; (5) deadline tracking with Level 1 and Level 2 escalation; (6) read-only AI analyst, Data Explorer, forecast, carbon reduction simulator; (7) water and waste tracking. Committed builds: (8) SHA-256 duplicate shield; (9) side-by-side verification screen; (10) six-pillar validation with anomaly detection; (11) statutory CEA/IPCC factors; (12) append-only audit log and ZIP audit pack; (13) carbon credit management; (14) internal carbon pricing; (15) cross-project monitoring and platform settings; (16) UI UX improvement (17) Responsive UI for both mobiles and desktop and fixing UI overflow
- Tech stack and hardware you plan to use
Frontend: Next.js 16, React 19, TypeScript, Tailwind CSS v4, shadcn/ui, Radix UI, Lucide React. Backend and APIs: Node.js, Next.js Server Actions, API Route Handlers, tRPC, Zod. Database and Authentication: PostgreSQL, Supabase, Supabase Auth, Row-Level Security (RLS), role-based access control. Document Storage and Reporting: Supabase Storage, SHA-256 evidence verification, pdf-lib, Puppeteer, Chromium, ExcelJS, SheetJS, JSZip. OCR and Document Processing: Python 3.12, PaddleOCR 3, PP-OCRv5 Mobile models, PaddlePaddle, OpenCV, pypdfium2. Emissions Engine: TypeScript-based deterministic calculation engine, versioned emission factors, unit conversion, calculation traceability. AI: NVIDIA NIM API with an OpenAI-compatible interface and restricted read-only analytical access. Testing and Deployment Tools: Node.js native test runner, Python unittest, pgTAP, Git, GitHub Actions, Docker, Vercel (web deployment target), Render (OCR worker deployment target). Hardware: No dedicated hardware is required. A computer with a modern web browser and an internet connection is sufficient for the demonstration.
- What have you built or tested so far?
We have built a functional prototype of EmeilSG, an evidence-first ESG and BRSR reporting portal. 1. BRSR Reporting: All 146 questions across Sections A, B and C mapped to the July 2023 SEBI BRSR format. 2. Organization Management: Group, company, subsidiary, business unit and project hierarchy with role-based access and segregation of duties. 3. Document Processing: Document uploads, PaddleOCR, PP-OCRv5 Mobile, 47 activity profiles, 20 templates and multi-bill PDF splitting. 4. Validation: Six validation rule categories, duplicate detection, consistency checks, anomaly flags and approval workflows. 5. Multi-Entity Consolidation: Group-level aggregation, Scope 1 and 2 drill-down and financial-year roll-ups. 6. Scope 1 Emissions: Supported fuel calculations covering diesel, petrol, CNG, LPG, coal, LNG and PNG, plus process and fugitive emissions. 7. Scope 2 Emissions: Electricity-bill processing, configured Indian utility profiles, and applicable grid- and market-based accounting. 8. Scope 3 Emissions: Selected categories covering waste, transportation, travel and capital materials such as cement and steel. 9. Calculation Engine: Deterministic CO₂e calculations, unit conversion, versioned emission factors and traceability. 10. ESG Analytics: Multi-period comparisons, emissions breakdowns, intensity metrics, hotspot ranking and peer benchmarking. 11. AI ESG Analyst: NVIDIA NIM-powered, restricted read-only analysis with eight guarded SQL tools. 12. Narrative Drafting: Seeded BRSR questions, 12 auto-filled tables, versioned disclosures and citation-supported drafts. 13. Carbon Reduction Simulator: Tests hypothetical emissions-reduction scenarios without modifying official records. 14. Reporting: PDF generation, disclosure management, dashboards and structured BRSR reporting. 15. Audit Evidence: Private evidence storage, SHA-256 verification, downloadable evidence ZIP bundles and audit trails. 16. Deadline Management: Reporting deadlines, reminders.
- Pitch deck (PDF)
- Pitch deck (PDF)PDF
- GitHub repository
- github.com/Dineshkumarpatra/EmeilSG
- What's working now?
The system has successfully integrated the core functionality, demonstrating reliable operation across critical modules. The data flow architecture now handles inputs efficiently,delivering consistent and predictable outputs. Ongoing efforts concentrate on refining performance and expanding support for enhanced features. EmeilSG is a fully functioning, audit-grade platform that turns raw operational bills into official SEBI BRSR reports with manual form-filling and OCR. The system ingests utility bills, fuel invoices, and waste manifests in any standard format, intelligently splits multi-page batch documents into individual monthly statements, and reads them locally through an AI extraction engine with built-in confidence checks. A mandatory side-by-side verification screen and a six-category automated validation engine catch duplicate invoices, out-of-period dates, and historical consumption anomalies before any numbers are finalized. Validated entries are calculated using deterministic mathematical formulas and versioned emission factors across Scope 1, Scope 2, and Scope 3, leaving a permanent, tamper-evident trace for every figure. To satisfy regulatory assurance requirements, the platform enforces a strict four-eyes principle where employees cannot approve their own submissions. Once approved, figures roll up automatically through project sites, business units, and subsidiaries into the group holding company for any custom period or financial year. Executives and department leads have dedicated role-based workspaces, real-time submission tracking that escalates overdue reports up the management chain, and a Decision Center featuring Tracey AI for conversational queries, an interactive What- If simulator to model decarbonization projects, and a carbon offset procurement estimator. Every reported number can be traced directly back to its source document, fully secured by immutable audit logs, strict data isolation.
- Demo video
- youtu.be/H9_Y06kvrPY
- GitHub repository
- github.com/IAMUCN/ps08_BRSR_meil.git
- Deck or report (PDF)
- Deck or report (PDF)PDF
AI tools used
Use ChatGPT for the prompting, the context storing, and the progress tracking; Claude for the implementation; Gemini for the minor UI and UX changes. Used Claude and Gemini to structure the 146 SEBI BRSR disclosure schemas, build deterministic carbon calculation formulas, and generate automated validation test suites.