3. Two strangers, same search, now on a call together (illustration of the call step).
5. Privacy first: activity stays on your laptop. Llama summarizes it locally, and only that short summary is sent.
4. You choose who you match with by age, country and language. Both people's filters must allow it.
1. A real match from work-match: two people working on the same thing get matched automatically.
2. Match, call, keep in touch: both say yes, a video call opens, then you can stay work friends.

Inspiration - Working alone is isolating. Somewhere, someone is working on the same thing as you right now, but finding them means posting in a Discord or going to a meetup, so most people just grind alone. We wanted that connection to happen on its own, at the moment you need it.

What it does - Your desktop notices what you're working on and finds someone doing it too.

TaskMatch sits in your menu bar and picks up what you're on, like a LeetCode problem, job applications or a CS course. When someone else is working on the same thing right now, you get one low-pressure popup: "Priya is also doing LeetCode right now. Want to work alongside her?" If you both accept, a video call opens. You can add each other as connections and get notified when they're working on that topic again. You do nothing extra.

How we built it - TaskMatch is a Tauri desktop app (Rust and JavaScript) for macOS and Windows. A Chrome extension and an app tracker capture what you're doing, private pages and apps are filtered out, and Llama 3.2 runs on your laptop to turn the rest into one short sentence and a topic. Only that summary is sent to our FastAPI and Supabase backend, where Jev, a decision model, scores how well your task matches everyone working right now. Calls run on Jitsi.

Challenges we ran into - Getting a small local model to label topics consistently was hard. It kept naming the website instead of the activity, like NeetCode instead of LeetCode. Deciding what counts as a match was harder. Someone coding Dijkstra's algorithm and someone watching a shortest-path lecture should match, but two people writing cover letters for different fields shouldn't.

Accomplishments that we're proud of - We built an eval of 137 hand-labelled pairs and used it to take matching accuracy from 58% to 96%. Your raw activity never leaves your laptop. And it's a real desktop app on two platforms that installs with one line.

What we learned - Evals beat vibes. Every time something "felt better," the eval told us whether it actually was. We also learned that small local models do great work when given a narrow job.

What's next for TaskMatch - Calendar awareness, group sessions, Discord calls, and screen reading for apps beyond the browser.