Intert AI

Inertia holds the clocks autistic users can't feel while absorbed — water, food, deadlines — and interrupts the same cheap way every time. Your clocks, your laptop, nothing leaves it.

Comment

Inspiration

I have ADHD, and when I hyperfocus I forget to eat, drink, or get up — I just don't feel it happening until it's already a problem. Soham and I realized autistic people describe the same gap under a different name: autistic inertia. Every productivity app out there helps you start something. None of them help you stop. So we built the missing half.

What it does

Inertia runs quietly on your laptop and holds the clocks you can't feel while you're deep in something — water, food, stretching, deadlines — all set in your own words, on your own timing. It stays silent until it decides an interruption is actually worth what it costs you: a small card first, escalating to full-screen only if you keep ignoring it. Every "not now" and every "okay" teaches it something, so it learns over time which moments to protect and which it shouldn't.

How we built it

It's a local-only Flask app — no cloud, no account, nothing sent anywhere. Two systems run separately: a pressure model that's fully deterministic (how overdue a need is, from the interval you set), and a disruption model that's learned (whether right now is actually a bad time to interrupt you). The disruption side uses a hierarchical Bayesian model with Thompson sampling, so it can make a reasonable guess from a handful of ratings instead of overfitting to noise.

The project is grounded in a real study — Rapaport, Clapham, Adams, Lawson, Porayska-Pomsta & Pellicano's 2023 interview study of 24 autistic adults on the experience of starting and stopping tasks: "I live in extremes": A qualitative investigation of Autistic adults' experiences of inertial rest and motion, published in Autism, 28(5), 1305–1315 (doi: 10.1177/13623613231198916). Nearly every mechanism in the app traces back to a specific finding in that paper.

Challenges we ran into

We tested Inertia with an autistic friend of Soham's, and on day two it interrupted him right as he was about to finish something — the app could tell he was absorbed but had no way to know he was 90 seconds from done. That's the exact harm we set out to prevent, and we'd caused it. We added an "almost done" option so he could tell the app what it couldn't sense on its own. After the fix, he told us it genuinely helped him — which mattered more to us than any other part of this project.

Accomplishments that we're proud of

Getting personalization to work on almost no data — the hierarchical posterior means a new context starts with a reasonable estimate instead of a blind coin flip, and narrows as real ratings come in, with no training step required. In a 21-day simulated benchmark, the learned policy beat the best non-adaptive baseline by 14%, with fewer interruptions landing well more often. But what we're proudest of is simpler: a real autistic person used it and said it helped.

What we learned

You can't remove the interruption — inertia means stopping always costs something. What you can do is make it predictable, so it stops being a surprise and becomes something the person already knows how to handle.

What's next for Inertia

Packaging it as a single-file executable so there's no Python setup at all. Testing with more autistic users beyond the one friend who tried it, so the personalization isn't shaped by just one person's patterns. And refining seam detection so more natural pauses can absorb an interruption for free.

Built With

Share this project:

Updates