Develense

Inspiration

Assignment instructions at school are almost never written for how a neurodivergent brain actually processes information. A single Google Classroom post can bury the actual task inside three paragraphs of context, a linked video, and a rubric PDF and for a student with ADHD, that's often exactly where the stall starts, way before the "hard part" of the assignment even begins. Tools like Goblin Tools already break tasks into steps, but they're generic. They don't know the student, they don't know the actual assignment, and they don't get better with use. We wanted something that lived inside a student's real school context and actually adapted to them.

What it does

IncludAI is an assignment copilot for neurodivergent middle and high schoolers. A student feeds it a real assignment a doc, a linked video, even a voice note from a teacher and it breaks it down into small, manageable steps. Instead of starting from scratch every time, it tracks where the student tends to stall week over week and adjusts how it breaks things down and how much support it gives, so the help gets more personal the longer it's used.

How we built it

We built on Next.js, Supabase, and Vercel, with the Claude/GPT APIs doing the core breakdown and adaptation logic. The ingestion side ended up being the hardest engineering problem — assignments show up in every format imaginable, so we used MarkItDown to normalize most documents, built a custom pipeline to pull and transcribe video instructions off YouTube links, and added audio transcription for voice notes, since a lot of real instructions come in as a recording, not clean text.

None of us had a direct line to a neurodivergent K-12 student going in, so early on we leaned on research instead of firsthand testing. Partway through, we connected with a neurodivergent person in Madras who tried the tool and gave us real feedback, and that's what actually shaped how we tuned the adaptation logic.

Challenges we ran into

Time. Between the ingestion pipeline, the adaptation logic, and the frontend, there wasn't enough runway to do all three justice — we ended the hackathon with a working demo, not a deployed product.
The adaptation logic. Getting the AI to actually learn how a specific student stalls, instead of just re-running the same generic breakdown every time, was a much harder problem than we expected going in.
Finding someone to test with. You can't just arrange access to a neurodivergent student or teacher overnight, so most of our early calls were made on research alone, and real validation only came later.

What we learned

"Personalization" is a lot easier to say than to build. Going from a generic task-breakdown tool to one that actually adapts to a person takes real signal over time, not just a clever prompt. We also underestimated how much of a project like this is ingestion wrangling messy, multi-format input rather than the "interesting" AI part everyone pictures. But the biggest thing: talking to even one real user changed our thinking more than all our research combined. That's what makes us want to get IncludAI in front of more actual students and teachers next.