Back to Biswaranjan's profile
13views

MEIL OneSource - One Central Source for Corporate ESG & BRSR Information

Talk to your ESG report.

Biswaranjan Sahoo
  • Video of it working
  • Problem, approach and result explained
  • Public source code
  • Expert reviewed· 6 reviews
MEIL OneSource - One Central Source for Corporate ESG & BRSR Information

80/80

BRSR indicators fillable by voice (up from 24)

~3×

Faster AI voice commands (2–3.5 s → ~1 s)

166

Automated tests passing

Walkthrough

Overview

The problem

Large Indian companies now have to report ESG data in SEBI's BRSR format, and BRSR Core asks for figures that can be checked against evidence. For a group like MEIL, that data starts on hundreds of project sites: diesel in a generator logbook, electricity on a bill, water from a tanker slip, injuries in a safety register. Today it travels up through spreadsheets and email to the business unit, the subsidiary and the group ESG team. By the time it reaches a report, nobody can easily say who entered a number, which bill proves it, or who approved it.

The people who enter the data are site engineers, EHS officers and store staff, not ESG experts. They work on sites with weak internet, often with their hands busy.

We set out to build one place where each person sees only what they're responsible for, enters it with proof, and sends it up a clear approval chain. The result should be BRSR figures the group can stand behind, with a trail from the report back to each bill.

The approach

Stack:

React + Vite front end, Express API and PostgreSQL (Drizzle), in a TypeScript monorepo. A shared package holds the BRSR indicator catalog and the approval rules, so the browser and the server always agree on who may enter or approve what. How it fits together:

Roles decide everything. There are 15+ roles, from site engineer to board. One access table drives the menu, the pages and the voice assistant, and the server checks the same rules again. Approval chain: Project → BU → Subsidiary → Group → Auditor → Board. Review starts at the level where data was entered, so Group HQ entries don't wait for a project manager who doesn't exist. Reviewers get an email when it's their turn. Built for sites: The app opens offline, and saves are queued and synced when the connection returns. Bills (PDF or photo) are attached to each value. Sign-in uses an emailed code and an optional face check. The tricky part: a voice assistant that runs on a small local model. We wanted hands-free entry ("generator 120 and petrol 40, then save") without sending company data to a cloud AI. Our model is Qwen 2.5 3B on a 4 GB GPU. Sending every command to it was too slow (2–3.5 s), and it sometimes filled the wrong fields. So we gave it as small a job as possible:

The browser first searches the real indicator list and handles clear commands itself, in milliseconds. That search tolerates misheard words ("dizel", "light bill"). Only unclear commands go to Qwen, with the 2–6 likely fields. The sentence is split into one part per spoken value, and Qwen must answer exactly one field and value per part. Every answer is checked: the field must belong to your role, and the number must actually have been said. If not, or if the model is slow, the assistant asks "Did you mean…? yes or no" instead of guessing. Left out on purpose:

No cloud AI or cloud storage for data: MEIL's figures stay on its own machines. The AI never writes on its own authority: every value goes through the same checks and approvals as manual entry.

The result

Working end to end:

sign-in with an emailed code; role-based data entry with evidence; the full approval chain with email alerts; BRSR readiness, calculator and reports; a supplier register with Excel import and the top-75%-of-spend boundary; offline mode; reminder emails every 3 days plus deadline alerts at 30, 14, 7, 3 and 1 days; the voice assistant. Voice assistant, before and after our redesign:

Before After Indicators reachable by voice 24 of 80 80 of 80 Clear commands ("diesel 450") sent to the model in the browser, near-instant Unclear commands via Qwen 2–3.5 s 0.65–1.3 s Wrong extra fields filled by the model yes, in our tests no: one answer per spoken value Locked entries (already submitted) showed "saved" when the save had failed refuses with the reason All 166 automated tests pass. In a real browser, the assistant handled single commands, multi-step commands, unclear words, locked entries and pages the role can't open, each with the right outcome.

How it was measured: times are wall-clock from the API on our laptop's RTX 3050 (4 GB) with the model loaded. Indicator counts come from the catalog. The browser checks were scripted runs in headless Chrome with commands typed into the voice panel.

Reflection

What's unfinished or known-broken:

Logins and company data are demo data, and sign-in still uses development shortcuts (a shared demo password). These must be replaced before real use. We tested voice with typed commands, not yet with many real voices on a noisy site. In Chrome, speech-to-text still goes through the browser's own speech service. The local model can be pushed out of GPU memory by other work. We added an automatic reload, but a cold start still costs about 3 seconds, and then the assistant falls back to asking yes or no. What we'd do differently: build the indicator catalog and role rules as one shared source from day one. Most of our bugs were places where the screen, the server and the voice assistant disagreed about who owns a field or where review starts.

What we learned:

A small model is fast and reliable when it gets a narrow, checkable job ("which of these 3 fields, and what number?"). It isn't when it's asked to understand everything. The most useful AI behaviour was knowing when to ask instead of guess.

Submitted to The BRSR Reporting Portal

Your idea in one line

MEIL OneSource is a centralized ESG and BRSR reporting platform for Megha Engineering & Infrastructures Limited (MEIL) that enables project teams and business units to record sustainability data, calculate greenhouse gas emissions, attach supporting evidence, and consolidate verified information into traceable, group-level sustainability reports.

What will be working by the finale?

By the Grand Finale, our goal is to deliver a stable, integrated MVP of MEIL OneSource that demonstrates how ESG data moves from individual project sites to consolidated group-level BRSR reporting. Our key commitments are: 1. Complete End-to-End Demonstration: Demonstrate a complete reporting journey, from entering project-level activity data to final review, approval, and consolidated reporting. 2. Reliable Calculations and Consolidation: Verify that sample fuel and electricity consumption data produces consistent emissions calculations and that approved records contribute correctly to higher-level reporting. 3. Functional Governance Workflow: Demonstrate role-specific access, all five approval gates, rejection and correction, resubmission, and final record locking. 4. Verifiable Data and Evidence: Show how a reported value connects to its supporting document, approval history, and audit trail. 5. Reporting-Ready Dashboard: Present consolidated ESG indicators, BRSR Core information, reporting completeness, and available report-generation or export functionality. 6. Demo-Ready Deployment: Prepare seeded demonstration data, role-based test accounts, and a stable application environment for the jury. Our priority is to demonstrate a reliable, traceable reporting process with consistent data and calculations, rather than simply presenting individual screens or unfinished features.

Tech stack and hardware you plan to use

Frontend: React 18, Vite, and TypeScript for the web portal, dynamic forms, dashboards, and reporting interfaces. Backend: Node.js and TypeScript with a REST API using the Fastify/Express framework configured in the project. Database and ORM: PostgreSQL 17 with Drizzle ORM for structured ESG data, organizational hierarchies, user roles, reporting periods, calculations, approvals, and audit records. Calculation engine: decimal.js for precision-based greenhouse gas and emissions-intensity calculations. Project structure and development: Turborepo and pnpm monorepo with shared packages for database access, calculation logic, and common types. Database security and auditing: PostgreSQL triggers for audit logging and protection of audit records and approved data. Containerization: Docker Compose for the local development environment and PostgreSQL setup. Hardware: No specialized hardware, sensors, or embedded boards are required. The application is software-based and will run on development laptops using a local PostgreSQL database and web browser. Remaining requirements: Final testing of the end-to-end workflow, verification of role permissions and calculation outputs, preparation of demo data and supporting documents, and confirmation that the application and database run reliably in the final demonstration environment.

What have you built or tested so far?

We have established the core application and technical foundation of MEIL OneSource, our ESG and BRSR reporting platform for MEIL. Work completed so far includes: 1. Application Architecture: Developed the React, Vite, and TypeScript frontend structure, a Node.js/TypeScript API, and a PostgreSQL database layer using Drizzle ORM within a pnpm monorepo. 2. Database and Demo Setup: Defined the core data models for organizational hierarchies, users and roles, reporting periods, ESG activity data, calculation results, evidence, approvals, and audit records. The repository also includes database seeding for demo accounts, organizational data, and emission factors. 3. Role and Organization Model: Established the role-based access structure and hierarchy connecting group, subsidiary, business-unit, and project-level reporting. 4. Calculation Logic: Implemented a decimal.js-based calculation module for greenhouse gas emissions and emissions-intensity metrics, with configured emission factors. 5. Reporting and Review Interfaces: Developed the application structure for data entry, review and approval, audit inspection, and BRSR Core reporting. 6. Security and Audit Mechanisms: Added database trigger definitions intended to record changes and protect audit records and approved data. 7. Evaluation Preparation: Documented a step-by-step verification workflow covering demo login, sample fuel-data entry, emissions calculation, evidence submission, sequential approval, and audit-trail inspection. We are focusing next on validating these components together, resolving integration issues, and confirming that the complete workflow functions reliably in the demonstration environment.

Pitch deck (PDF)
Pitch deck (PDF)PDF
What's working now?

MEIL OneSource is a working web app for collecting ESG data across MEIL's sites and companies and turning it into SEBI BRSR reports. Working today, end to end, on our local setup: Role-based sign-in for every MEIL role (site engineer, EHS, HR, store, project manager, BU, subsidiary and group ESG heads, auditor, board, IT admin). Each role sees only its own pages and tasks. Two-step sign-in and password reset with a 6-digit code sent by email, plus optional face check. Data entry for 80 BRSR indicators: values, supporting bills (PDF or photo, up to 15 MB), and AI reading of uploaded bills. Review chain: Project Manager → BU → Subsidiary → Group ESG → Auditor → Board. Review starts at the level where the data was entered, and each reviewer gets an email when it's their turn. Reports: BRSR readiness, BRSR Core calculator, reports, traceability and audit trail pages. Value chain (suppliers): a Group-level supplier register with Excel import and preview, and the top-75%-of-spend boundary worked out automatically. Works offline: the site opens without internet, saves are queued, and they sync when the connection returns. Email reminders: pending work every 3 days, and filing-deadline alerts at 30, 14, 7, 3 and 1 days. IT admins configure email settings in the app. OneSource voice assistant: say "diesel 450" or "generator 120 and petrol 40 then save". Clear commands run instantly in the browser. Unclear ones are settled by a local Qwen 2.5 (3B) model in about 1 second. If it's unsure, it asks yes or no.

Deck or report (PDF)
Deck or report (PDF)PDF
A photo or screenshot
A photo or screenshot: MEIL OneSource - One Central Source for Corporate ESG & BRSR Information

AI tools used

Gemini

API Integration, Backend development

Links & files

Artifacts

1

Gallery

5