Six months in Silchar — what running AI literacy taught me.
Before I wrote a line of product code, I spent six months in Silchar, Assam. The plan was to run on-the-ground AI literacy work in CBSE schools and learn how Indian schools actually decide. I came back with three things I had wrong in the first month, and two convictions that became Deshika's spine.
I had the wrong target. The first month I spent talking to principals, because principals sign the cheques. That was correct and incomplete. Principals listen to trustees, in the way that a CEO listens to a board, and trustees in most affordable-private and Tier-2 schools care about three things: parent satisfaction, board-exam results, and whether the school's name will appear in a regional newspaper for the wrong reason. AI conversations in 2026 land in that third category by default. If you want the principal's attention, your pitch has to give the trustee a defensible position to a parent WhatsApp group in November when the first incident hits a peer school. That changes the opening line of every conversation.
I had the wrong reader. The second month I drafted what I thought was a good parent communication on AI use in the school. Two pages of English, clean prose, the kind of thing I would have sent in Bangalore. The school's academic coordinator read it, complimented it, and then explained that thirty percent of the mothers in the parent WhatsApp group couldn't read it. The communication that actually worked was an eight-minute voice note in Bengali (Silchar is Bengali-speaking, not Assamese, even though the state is Assam) that the coordinator recorded herself, sent on a Sunday evening, and asked mothers to share with husbands. That is what got read. Or rather, that is what got listened to. The parent-facing document for the affordable-private and govt-school clusters in India is not a document. It is a voice note in the regional language.
I had the wrong unit. The third month I tried to measure adoption per-student. That number told me nothing. The unit that mattered turned out to be per-section. A Class 9 section is twenty-eight kids, one class teacher, one set of parents who all know each other, one head of department who reports to one academic coordinator. Adoption travels through that section as a unit or not at all. The first section that goes well in a school produces the second one. The first school in a Sahodaya cluster that goes well produces the next three. If you optimise for student-by-student you build the wrong product. If you optimise for section-by-section you build the right one.
Two convictions came home with me. The first: every Class 8 student in every CBSE school is already using free ChatGPT, and the class teacher knows it. The conversation in the staffroom is not whether — it is what to do about it. That changes what "AI in school" means. It is not a future-readiness project. It is a present-reality intervention. The second: the DPDP Act, which most schools have not yet read, gives those schools the legal foothold to do something about that present reality — to authorise a tool they can stand behind and to disauthorise the tools they cannot. The combination is the whole product thesis.
The Silchar months produced one more thing that does not show up in any deck. The way you find out whether you have built something useful is to drive to a school on a Wednesday, sit in the back of a Class 9 section for forty minutes, and watch what happens when the AI gives a wrong answer. The student notices. The teacher notices. The product either survives that moment or doesn't. Six months of those Wednesday afternoons are what Deshika is built on. That is the difference between desk research and what the brain calls the load-bearing stuff.