This blog is written by Ashwin from Madhi Foundation.
From a Chatbot to a Teacher’s “Second Brain”: Our Journey Through the AI Cohort
For a long time, Gokul and I have been pondering over how to make VallamAI more than just a chatbot, how to make it a tool that could actually aid teachers in reaching every child effectively. But our daily work never really gave us the space to chase that. That changed when Madhi got selected for the AI Cohort by Project Tech4Dev. It gave us a space to reflect and think deeper, and we were assigned a mentor, Vinod, who kept questioning our approach and kept us from over engineering things we didn’t fully understand yet.
After the first cohort session with Rishikesh, we thought we’d walk away with clarity and just get started. Instead we came back with more questions than answers, questions that genuinely challenged us. So we went straight to the ground, sitting with teachers, listening rather than assuming what they’d want from VallamAI. In March 2026 we spoke with the teachers across Tamil Nadu, and the same story kept surfacing. One teacher put it simply: “They don’t go back to scores, they teach from the levels they already carry in their heads.” Across the interviews, 77% told us they rely on what they observe in class rather than formal assessment scores, 86% could already spot which child was struggling and 70% said they’d rather assess a child by listening to them than by marking a paper. That told us something important. The knowledge that actually drives good teaching lives entirely in the teacher’s head, and it never makes it into any system. That’s the gap we realized VallamAI needed to close.
That set us on the path of building what we now think of as a second brain for the teacher. A system that generates dynamic lesson plans, contextual to each classroom, pulled from a context generator engine that brings together classroom context, teacher observations, student performance, and a daily journal the teacher fills in themselves. The journal is supported by a voice recording feature, so it works like a digital diary where teachers can record their thoughts in their own native language and flag specific students who might need more support for the next lesson. The lesson plan generator then takes all of this and builds a plan that stays true to the existing handbook structure, but layers in contextualization specific to that classroom, with remediation strategies the teacher can lean on to actually move the needle on learning outcomes.
The cohort kept pushing our thinking in ways we didn’t expect, especially the Tattle session on guardrails. It pushed us to immediately try and jailbreak our own chatbot, and that’s where we ran into our next real problem: how do we untie the darkness sitting inside the AI’s black box. Since we work at scale, a fully transparent system isn’t optional for us, we need to know exactly how every lesson plan gets stitched together at every level. That search led us to Langfuse, which let us build observability at a very granular level, giving us visibility into every layer of the system and showing us precisely where things needed to change.
We’ve since taken this out of the lab and into real classrooms. We ran it past a group of 10 teacher advisors, and the feedback was honest. What landed well: it genuinely saves prep time, the reteaching strategies for the whole class felt concrete, and it met teachers where they were while still holding on to the THB structure. What needs work: the flow and pacing of activities, sharper remediation suggestions for smaller groups, and better contextualized materials for each activity. The numbers backed up what we were hearing too, prep work that used to take close to 8 hours a week is coming down toward 30 minutes, at a cost of around INR 2 per student per month, covering the voice journals, lesson plans, assessments, and the chatbot together.
What we’re learning along the way matters just as much as what we’re building. We’re constantly finding the balance between contextualizing a lesson plan enough to be useful and over engineering it into something unwieldy. We’re learning that teachers bring very different skillsets and experiences to the classroom, and supporting each of them in a way that plays to their own strengths is going to matter more than we initially thought. And maybe the hardest part, we’re learning that reflection isn’t a habit most teachers already have. Building that habit, in a way that genuinely shows them it helps, might end up being our biggest win of all.
Good tech gets out of the way. Great tech leaves something behind. We’re aiming for the latter.
