India's tinkering labs get an AI teaching assistant, powered by Gemini

Key points

• Google DeepMind and India's Atal Innovation Mission launched ATL Saathi on July 14, 2026, a Gemini-based web app for teachers in Atal Tinkering Labs.
• The pilot covers 100 schools, part of a network that reaches more than 1.1 crore students across India.
• The assistant generates grade-appropriate project ideas, step-by-step assembly guides, wiring diagrams and safety precautions.
• It launches in eight Indian languages, with room to add more.
• It runs on Gemini 3.5 Flash, Google's faster, lower-cost model tuned for quick responses.

India has spent years building Atal Tinkering Labs, small school workshops kitted out with 3D printers and basic robotics kits. There are enough of them now to reach over 1.1 crore students. The hard part was never the hardware. It was finding teachers who could confidently walk a curious 13-year-old through wiring up a soil-moisture sensor without setting the classroom on fire.

That's the gap Google DeepMind is trying to plug with ATL Saathi, a web app for teachers that launched in pilot on July 14, 2026. The day before, we reported on Anthropic's move to introduce rupee pricing for Claude in India, a sign of how quickly the country is becoming a front line for AI product launches.

Think of ATL Saathi as a patient co-teacher that never sleeps. A teacher opens the app, picks a module from the official tinkering curriculum, and gets a short summary, an infographic and a quick quiz. No lengthy training video. No PDF running to hundreds of pages.

The more interesting part is what happens when students turn up with an idea.

Say a student wants to build a device that warns a farmer when a water tank is nearly empty. The teacher types the problem into ATL Saathi. The app, running on Gemini 3.5 Flash (Google's fast, low-cost version of its main AI model), returns a project plan with the relevant components, a wiring diagram, assembly steps and safety notes about loose batteries and exposed wires.

It works the other way too. If a teacher wants to spark ideas rather than respond to them, the app generates project suggestions matched to the student's grade and the current syllabus. Eight Indian languages are supported at launch, so a teacher in a Tamil-medium school in rural Tamil Nadu gets the same wiring diagram and the same safety warnings as a teacher in an English-medium school in Delhi. The curriculum materials sit inside NotebookLM, Google's document-grounded AI tool, so the assistant draws from the official Atal Innovation Mission playbook rather than the open internet.

The Atal Innovation Mission sits inside NITI Aayog, the Indian government's policy think tank, whose stated goal is to cultivate one million Indian children as what it calls "neoteric innovators". The partnership was announced at the AI Impact Summit in February 2026, and Google DeepMind describes today's launch as the first working product from that commitment.

What does this actually change for a classroom?

It shifts the teacher's job from hunting for information to guiding a student through it. A physics teacher who has never touched an Arduino board, the small programmable circuit board students use for electronics projects, can walk into a lab with a plan and a parts list generated in minutes.

A specialist mentor visiting each of the thousands of tinkering labs across India isn't affordable at national scale. Even one visit per school per term would run into serious money and logistics. A software assistant that costs pennies per query, running on a model designed for cheap high-volume use, is a very different economics problem.

There are real questions the pilot will need to answer. Whether the wiring diagrams hold up when actually built, and whether the model ever invents a component that isn't in the lab's kit, are things only classroom use will reveal. The 100-school cohort is where those answers come from. If Google DeepMind and the Atal Innovation Mission can show that teachers spend less time on paperwork and more time supervising actual builds, scaling to the full network becomes a much easier case to make.

For now, the tinkering labs have a new staff member. It answers at 2am, speaks eight languages and doesn't tire of the same question about resistors.