Inspiration We kept noticing the same thing over and over. Someone struggling on a website, not because the content was bad, but because the page itself was working against them. Too much clutter, text too small, some flashing banner ad making it impossible to focus. And when we looked into "fixing" it, the answer was always the same: dig through ten menus of browser settings you didn't know existed, with names that don't mean anything to a normal person. We wanted something where you could just say what's bugging you, in your own words, and have it actually get fixed.
What it does Comis is a browser extension that lets you describe what's wrong with a page, typed or spoken, and fixes it on the spot. Say "this is too cluttered" and it strips out ads, menus, and pop-ups. Say "make it easier to read" and it adjusts text size, contrast, or turns on a cleaner layout. It can simplify pages, reduce motion, read content aloud, reposition layouts, and more, all instantly, all undoable with one click, and all without sending your data anywhere.
How we built it Under the hood, most commands are handled by a local parser so common requests work instantly, even offline. When something's genuinely ambiguous, we fall back to AI (Gemini), but only to interpret intent, never to generate or run arbitrary code. Every AI response gets validated against a strict, pre-approved set of safe instructions before it touches the page. If the system isn't confident it understood you, it asks instead of guessing.
Challenges we ran into Getting the balance right between "smart enough to understand vague requests" and "safe enough to never do something unexpected" was harder than we expected. We also had to make sure commands didn't stack weirdly if someone ran them twice, and that reset actually cleaned up everything the extension touched, no leftover styles, no half-broken pages. Building something that had to survive flaky Wi-Fi and still feel reliable during a live demo pushed us to rethink a lot of our error handling.
Accomplishments that we're proud of We're proud that Comis never feels like a black box. Every action comes with a clear explanation of what it understood and why. We built real privacy in from the start instead of bolting it on later. And we didn't fake any of our numbers. We built an actual evaluation suite so we could show real accuracy and speed instead of just claiming it works.
What we learned We learned a lot about the tradeoffs between local logic and AI, how much you can solve with simple pattern matching before you actually need a model, and how important it is to gate AI usage tightly so it can't misbehave. We also learned how much trust matters in accessibility tools specifically. People need to know a tool won't break their browsing experience before they'll actually use it.
What's next for Comis Next, we want to expand the range of commands it understands, add more nuanced reading modes, and improve the confidence gate so it asks smarter clarifying questions. We're also interested in exploring how Comis could learn a user's preferences over time, not by storing personal data, but by remembering the kinds of fixes they tend to want, so the experience gets faster the more you use it.
Log in or sign up for Devpost to join the conversation.