This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
The main quest can wait. One small escape at a time.
You finish work, you know there has to be more to life than another evening on a screen, and you still have no idea what to do with the next two hours. Sidequestar is for that moment: a quest-giver running on your own machine that hands you one small thing to do, and gets to know you a little better every time you do one.
You tell it your situation ("two hours this evening, want to be outside"), a model running on your own machine writes one specific quest, and you go do it. Finish it and you get XP and levels. Add a photo and you get a bit more. The whole thing on screen takes about twenty seconds, and then you're out the door.
Every time you finish a quest, the model rewrites a short summary of what you seem to enjoy and when you're free, and that summary shapes the next quest. I didn't want that to be a black box, so the profile page shows, word for word, what the model is told about you, and you can edit the summary if it gets something wrong. Everyone on the machine has their own account, so housemates get their own quests, XP and summaries.
Yesterday I went cycling in Tigoni with a group: 33.64 km, 3 hours 26 minutes and 532 m of climbing. That ride wasn't a Sidequestar quest, but it's the kind of thing I want more of, and its Strava card is the completion photo in the demo.
Demo
Try it: sidequestar.nyagah.me. Create an account and ask for a quest. It runs the same open model on a small 4-core ARM server, so a quest takes about 10 to 25 seconds. If you'd rather keep everything on your own machine, make dev runs the whole thing locally.
In the demo I create an account, ask for a quest on a Saturday morning with my bike and half a day free, finish it with a photo, and then open the profile page to see what the model knows. The waits for the model are sped up.
Code
Sidequestar
The main quest can wait. One small escape at a time.
Sidequestar gives you small things to do away from work and school, and XP for doing them. You describe your day, a model running on your own machine suggests one concrete quest, and you go do it. Over time it learns what you enjoy and writes that down where you can read it and correct it.
Built for the Hacktoberfest Open-Source AI Challenge, Week 1: Touch Grass. Try it at sidequestar.nyagah.me, or run it on your own machine (below) to keep everything there.
How it works
- Tell the quest-giver your situation, for example "two hours this evening, I want to be outside".
- A local model (Ollama running
llama3.2:3b) writes one specific quest. You can also add your own. - Go do it, then mark it done. Adding a photo earns bonus XP.
- XP adds up to…
How I Built It
The backend is Go (standard library net/http) with SQLite through sqlc. The frontend is Astro with Tailwind CSS and daisyUI. Quest generation and the profile summaries both run through Ollama with llama3.2:3b, an open-weight model, entirely on the user's machine. make dev starts Ollama, the API and the frontend together.
Most of the work went into getting a 3B model to behave. A few things I ran into:
- It made things up, and then kept them. Early summaries invented details like "free until 10am". Worse, each rewrite started from the previous summary, so once the model invented something, it treated it as fact from then on. The fix was to stop feeding it its own output. A rewrite now only sees what you wrote, the quests you actually finished, and your last manual edit.
- It treated starter quests as history. New accounts come with a few seed quests on the list, and the model read them as things the person had done ("She tends to finish short quests like reading or taking a 20-minute stretch", about a test user who had done neither). Labelling the list as "not done yet, not evidence" and dropping the temperature for summaries fixed most of it. It still slips now and then, which is why the summary is editable.
- It invented places. Early quests sent people to "Kiambu National Park" and "the Kiambu Tea Museum", neither of which exists, and this app's whole point is sending you out to find things. I tried
gemma2:9bto see if a bigger model would do better. On this laptop (an i5-8250U, no GPU) it took 88 and 115 seconds per quest, timed out on the third, and still sent a Limuru cyclist to "hidden art installations" at Ol Donyo Sabuk, about 80 km away. So the problem was my prompt: I'd asked for something "concrete and specific", and both models made specifics up. Now it describes the kind of place ("a tea farm near Limuru") unless you named one yourself. None of the five test quests I ran after that change named a made-up place. - The profile page reads its "what the model knows" lines from the same function that builds the prompt. I could have written a description of what gets sent, but it would have drifted out of date the first time I changed the prompt.
- If Ollama isn't running, or the model isn't pulled yet, the app says so and gives you the command to fix it, instead of "something went wrong". With a local model those are easy mistakes to make, and I made both while building this.
The design is a topographic map sheet: an orange map-cover header, contour lines behind the quest box, your active quests as a footpath between numbered waypoints, and finished ones as photo insets. Light and dark mode follow your system.
Why Does Open Innovation Matter?
Because of what this app knows about you. Your situation, your history and a running summary of who you are never leave your laptop. With a closed API, "I'm burnt out and have two free hours" goes to a vendor, every quest costs money, and the whole thing stops working wherever there's no signal, which is exactly where you want to be if the quest worked.
Ollama and an open-weight model also mean it costs nothing to run and isn't tied to a provider. Swap in another model, or replace Ollama with llama.cpp, and the code doesn't change.
Top comments (0)