Building Smarter, Not Just Faster: The Story of How AI Became Part of How We Work at SRM India

Foundations

It Started With Learning the Fundamentals

Before any of the tools, there was groundwork. Our early experiments with machine learning, including training custom models internally, taught us what AI could and couldn’t do reliably. That foundation mattered: when generative AI and large language models arrived, we weren’t starting from zero. We were ready to point them at real problems.

Accessibility

The First Real Problem: Accessibility

The first wall we hit was accessibility. Compliance standards, including WCAG, Section 508, and ADA, all come down to the same requirement: every image in a document needs a clear, accurate description. Simple enough, until you’re looking at thousands of images across a single title and thousands of titles across a year. Doing that by hand doesn’t scale. So we built PicScribe. The first version, Phase 1, used Azure AI Vision to generate alt text directly from the image itself. It worked, but it had a blind spot: it described what an image looked like, not what it meant in context, so it still needed heavy human review. Client feedback pushed us to Phase 2: contextual alt text. By feeding PicScribe the surrounding PDF content alongside the image, it could understand why an image was there, not just what was in it, cutting human correction effort dramatically. Today, PicScribe runs on Azure AI Foundry, built for enterprise-scale volume, with security and governance baked in from the start. That solved the image problem. But it also showed us something: once you start looking for places where manual review breaks down under volume, you keep finding more of them.

Epub Production

Then Came the EPUBs

The next one showed up in EPUB production. Foreign-language passages inside a document are easy for a human to miss, especially at scale, and they cause real problems for accessibility and rendering if they’re not tagged correctly. So we built an automated language identification tool that detects foreign-language content and applies the correct language attributes in batch across EPUB files, catching what manual review was always going to miss, simply because no reviewer can hold that much text in their head at once. It now processes 150+ titles a month.

Document Recovery

Then the PDFs That Wouldn’t Cooperate

Around the same time, we kept running into another familiar problem: source files that arrived as PDFs instead of editable formats, requiring painstaking manual reconstruction just to get back to a usable Word document. Using Azure Document Intelligence and OCR, we built a text extraction solution that preserves document structure while producing clean, editable Word output, reducing QA effort, improving accuracy, and speeding up migration work that used to eat entire timelines. It’s now processing 6,000 pages a month.

Engineering

The Bottleneck We Didn’t Expect: Our Own Engineering Time

By this point, AI wasn’t just fixing client-facing production problems — it was speeding up how we built the next tool, too. Our engineering team began using Claude to generate front-end components (React, Angular), back-end APIs (FastAPI, Django), and boilerplate code, leading to faster feature development, more consistent code quality, and less time on the repetitive parts. That acceleration is what made our next two platforms possible:

• • EditIQ — an AI-assisted content quality platform that automates validation against editorial standards, improving consistency and compliance while cutting manual review time.

• • SVEN — a browser-based, interactive learning authoring platform built as an alternative to Storyline, publishing LMS-ready HTML and SCORM content, with far simpler maintenance than what came before.

In other words, the tools were helping us build more tools.

Client Platforms

Taking It to Clients Directly

The same approach went straight into client work. For Core Knowledge Foundation, we built the Core Knowledge Media Asset Manager, a secure, cloud-based digital asset management platform with AI-driven metadata (auto tags, keywords, captions, alt text), bulk management, advanced search, and full life-cycle governance, with a Figma-designed UI and a Claude-built back end. It’s a good example of what happens when business teams, not just engineers, start driving AI-powered product ideas — something we’ve seen more of as agentic AI has made prototyping faster and more accessible to nontechnical teams. Alongside it, we’ve also been extending into video analysis and consolidating our growing toolkit into a centralized AI Tools Platform, a single hub for the AI services our teams rely on day to day.

Operations

And Finally, the Work Behind the Work

The last piece was about everything that happens around content: monitoring incoming requests and auto-creating job cards, auto-transferring client files into production, and auto-uploading final deliverables to client FTPs, OneDrive, or Dropbox on schedule. None of it is glamorous, all of it is necessary, and together it’s saving our teams 160+ hours a month, freeing our team to focus on client collaboration instead of routine setup.

Takeaway

Why This Matters

Every one of these tools started with a real production problem, not a strategy slide. AI at SRM India isn’t a pilot project sitting in a corner. It’s stitched into how the work actually gets done, day to day, across accessibility, production, engineering, and now client-facing platforms.