QED42

We worked for an academic health system to rebuild the public site for its College of Medicine and healthcare services. The site is where people find clinical trial details, search the physician directory, read the research newsroom, or open a patient-facing page. Before the rebuild, the site ran on a large legacy platform that held years of content. It could not all move at once, and a new launch date would not wait. Both usual options failed them. Migrate everything first and slip the date, or launch in phases and leave broken pages where the most important content used to be. So we separated the launch from the migration. Built a proxy inside Drupal that serves the not-yet-migrated pages live, inside the new design, so a visitor reaches the page either way and sees one consistent site. As each page moved into Drupal, the new version took over on its own. We scheduled no cutover and managed no redirects. The old site receded page by page as the migration went on. The new site went live on the day we promised, with every page working and the migration still running quietly behind it. Nothing broke, and no one could really tell what had already moved and what had not. The physician directory, clinical trials, and research newsroom now run as one connected ecosystem, publishing from a single content hub. The platform is the health system's to own and run for the long term. The complete case study 🔗 https://lnkd.in/d6sa5mjw

Video is the toughest data to handle. At the Oaisys AI Practitioners meetup, Pratik Bhande showed why and how he built a system that handles it. The reason video is hard is that it carries three kinds of information at once: the text on screen, the audio, and the visuals. And none of them stand alone. A transcript can't tell you what the speaker is pointing at. The visuals can't tell you what was said. You only get the full meaning when you combine all three. Traditional video RAG doesn't do that. It grabs one frame every second, so a 10-minute video becomes 600 frames, and most of them are duplicates because the screen barely changes. Newer systems tried to fix this by removing duplicate frames. But that created a new problem. Two slides can look almost identical while showing completely different text, so removing one throws away information. Pratik changed the question the system asks. Old systems ask how a frame changed from the last one. He asks whether any information changed at all. That idea comes from the attention paper. Not all tokens deserve equal attention, and the same is true for frames. So his system keeps only the frames that carry new information and drops the rest. On a 10-minute video, that takes 600 frames down to under 40. Ninety percent of the redundancy is gone, and none of the information is lost. The cost follows. Across 50 videos, the old approach costs around $120. His system costs around $22. One demo used a silent screen recording, with no audio and no narration. He asked it how to build a knowledge base in the tool, and it returned the exact steps, read entirely from the text and visuals on screen. Nothing was spoken. The system read it anyway. Watch the complete talk: https://lnkd.in/d_4JGXNm

At the Oaisys AI Practitioners meetup, summer edition 2026, Piyuesh Kumar gave a talk on the question most services orgs can't answer at the moment: what is the actual return on their AI spending? His starting observation was simple. Almost everyone now uses AI tools daily, but very few can quantify the benefit. People report feeling faster, or estimate a 2x or 5x gain, but those are impressions rather than measurements. And if the return is only a feeling, it can't justify the investment. His explanation for why the number is so hard to produce is the useful part. Tools alone don't create measurable value. A tool becomes valuable only when it's connected to how an organization works: its people, its existing knowledge, its standards. This is also why no off-the-shelf product can credibly measure your AI ROI. The measurement depends on your specific context. The approach he described is an operating system rather than a collection of tools, with people and organizational context at the center. It's organized around five pillars that build a framework created by QED42. Acquire: using your real signals, such as the technologies you specialize in and the certifications you hold, to find and qualify the right work. Deliver: speeding up the repeatable parts of delivery, like moving from design to working web pages, with quality checks built into the process. Operate: surfacing data that already exists in project tools, such as allocation, project health, and bench, into a single view so staffing and hiring decisions are informed rather than guessed. Prove: connecting AI coding sessions back to the underlying tickets, comparing actual time against the original estimate, and calculating the real efficiency gain, so a team lead can see the result in numbers. Adopt: feeding those signals back into coaching and reusable practices, and scaling what's working. He also reframed the common claim that AI costs more than humans at work. His point was that this is usually a measurement problem, not a cost problem. If the same team produces more in the same time, that's a real gain, provided you have the instrumentation to see it. The throughline: guidelines tell people what to do, but data tells you whether it worked and how to improve. The honest move is to stop estimating AI ROI and start measuring it. Watch the video on YouTube: https://lnkd.in/dhnpXHjr

We partnered with the American Diabetes Association to rebuild Diabetes Food Hub. Food Hub is where people with diabetes plan how and what they eat. They find recipes, set nutrition goals, build meal plans, and turn those plans into grocery lists. Before the rebuild, these tools worked separately. A recipe, a meal plan, and a grocery list did not share one source. We rebuilt Food Hub so they work as one. Users set their nutrition goals, plan meals against them, and see the nutrition update as the week takes shape. The plan becomes a grocery list, and the whole thing downloads as a clean document. Recipes, ingredients, plans, and lists now run from a single foundation, so what a user sees stays accurate across sessions and devices. The platform also carried over its full history. More than 263,000 returning users kept their saved recipes, meal plans, and grocery lists through the move. Diabetes is managed one meal at a time. Food Hub now holds that work in one place. Built on Drupal, the platform is ADA's to own and run, with full control of the foundation for the long term. Complete case study: https://lnkd.in/dTp9MU4j

⚡ Faster pages, better UX. At React Nexus 2026, Archana Agivale will dive into Partial Prerendering—a modern rendering strategy that improves Largest Contentful Paint (LCP) by combining the strengths of static and dynamic rendering. Learn how to ship blazing-fast experiences without sacrificing interactivity. 🎟️ Get your tickets now and join us at React Nexus 2026 on July 2–3. https://reactnexus.com/

Last week we opened the waitlist for EventHorizon. Today we are sharing a walkthrough of how it works. In the demo, Souvik Pal runs a real Drupal codebase through the suite from start to finish. Here is what you will see. The dependency map shows every module, service, and hook, and surfaces circular dependencies on its own. Selecting a function shows where it comes from, who calls it, and everything it touches. Each codebase gets one benchmarkable score, measured against industry standards for security, performance, architecture, and maintainability. The AI chatbot answers questions about the codebase. Asked which function carries the heaviest circular dependency, it returns the module, the function, and the line, with the source attached. Every analysis runs with no AI involved, and the codebase is never sent to a cloud model. When you want AI, you connect your own Gemini, OpenAI, or Anthropic key and choose the model. EventHorizon is built for teams that work with Drupal at scale: agencies inheriting client codebases, and enterprises planning upgrades and managing platform risk. Watch the demo: https://lnkd.in/gHCXEsxM Join the waitlist: https://lnkd.in/gSy9U4K2

We worked with Kotak Securities, now Kotak Neo, to rebuild their platform, including the way they publish IPO pages. An IPO is open for about three days. Investors use that short window to weigh the offer and decide whether to apply. An IPO page is only useful while the window is open. For Kotak, publishing one used to take up to two weeks. Each release had to pass through outside vendors and developer time before it could go live. We rebuilt the workflow with their team so they could publish directly. It now takes days, without outside vendors. The timing matters. An IPO page that is live while the window is open captures applications and new demat accounts that a late page would miss. For Kotak Neo, that is revenue. Complete case study: https://lnkd.in/dgnTGzPu

Today we're opening the waitlist for EventHorizon, a code-intelligence suite for Drupal. Point it at a repository, and it builds a complete picture of the codebase. You can explore, question, and share: architecture, dependencies, complexity, code health, caching, and config. The scanning engine already powers our open-source EventHorizon CLI. The suite grows from that same engine and opens it into a full visual platform. That platform is where the work happens: An interactive map of every module, service, and hook. Circular dependencies and upgrade blast radius show up at a glance, instead of buried in a report. Pick any function and see where it comes from, who calls it, and everything it touches. Find the parts of the codebase carrying too much, and the refactors that pay off. One benchmarkable score per codebase, with stats on functions, services, hooks, and dependencies. Technical enough for your leads, clear enough to hand a PM or client with no walkthrough. Performance and security analysis, written for how Drupal works: hooks, render arrays, routing, and caching layers. Every analysis works with no AI at all. When you want it, connect your own Gemini, OpenAI, or Anthropic key, choose the model, and get answers grounded in your project with source attribution. EventHorizon is built for the teams that carry the most Drupal at once. Agencies audit and inherit client codebases. Enterprises plan upgrades and manage platform risk. Audits that took weeks compress into hours, and the whole team sees the result, not just the person who ran the scan. Launches soon. Early-access teams get onboarding with our engineers, launch pricing, and a direct line into the roadmap. Join the waitlist: https://lnkd.in/grBjbUsv