Inspiration
I’m a third-year engineering student, and my university portal is where everything lives: quizzes, assignments, slides, room changes. The problem is how it’s posted. One week I found out my quiz date was hidden halfway through a paragraph about the classwork grade. No heading, no separate announcement. Every night I’d scroll through six courses asking “what do I have tomorrow?”, and I was always scared I’d missed something.
Studying was the same struggle. When I used an AI to prepare, I had to download and upload every lecture by hand just to get a decent answer, and it still didn’t know what my TA’s quizzes looked like. The material was already in the portal. I just wanted something that could read it for me.
What it does
DueHunter is a Chrome extension that sits next to your university portal.
One click syncs every course. It reads all the announcements, including weekly posts, and pulls out every graded deadline into one list sorted by date. It handles quizzes hidden mid-paragraph, postponed dates (it keeps only the new one), and different halls and times per tutorial group (it shows only yours). No guessing. Every deadline shows the exact sentence it came from. No uploads. It already has your lectures, practice sheets and sample quizzes from the portal. Readiness check. Tell it what a quiz covers (“lectures 7-9”) and it tracks how ready you are and whether you have enough study time left. Voice commands in English or Egyptian Arabic. “Remind me to study discrete on Thursday” adds it to your list. Smart search by meaning. “How do I turn Mealy into Moore?” finds the right lecture and summarizes it. TA-style mock quizzes generated from last semester’s sample quizzes, graded with a model answer and the slide it came from. A tutor that answers from your own slides and shows them. You can paste a screenshot of a question you’re stuck on and it explains every option. Add your own deadlines and PDFs for anything that isn’t on the portal. How we built it Chrome extension (Manifest V3) with a side panel. It reads the portal’s page structure directly, so lectures and files appear instantly with no AI needed. Firebase AI Logic in hybrid mode. Requests run on Gemini Nano on the laptop first (private, free, works offline) and fall back to Gemini in the cloud when needed. I added my own routing on top: Short, simple announcements and commands stay on Nano. Long, messy announcements, PDF summaries and photos go to the cloud. If Nano finds nothing, or only past dates, the cloud double-checks. pdf.js reads the lecture PDFs and renders slides as images, so the tutor can show the exact slide and the cloud model can read circuit and FSM diagrams. Firebase App Check protects the AI calls. Everything is stored locally (Chrome storage and IndexedDB), so uploaded files never leave the laptop. A public demo portal on GitHub Pages copies the real portal’s structure with fictional courses, so judges can try DueHunter without a university account. Challenges we ran into Nano missed the buried quiz. On long, confusing announcements the small model found the easy deadlines and skipped the hidden one. The fix was routing by length, plus a prompt that tells it events can be mentioned in passing. Wrong years. Nano sometimes wrote 2025 instead of 2026, so real deadlines ended up under “past”. I added a year fix and a cloud retry when every result is in the past. Hallucinated deadlines. The AI must quote the sentence each date came from, and anything it can’t back up with a real quote is dropped. Speed vs. rate limits. Syncing courses one by one was slow, and syncing all at once got rate-limited. I ended up loading pages in parallel, running at most 2 AI calls at a time, and caching results. Knowing which lecture a question is about. “Explain lecture 9” first opened the wrong course’s lecture. Now it matches lecture numbers and course codes before guessing by topic. Accomplishments that we’re proud of It found the quiz date hidden in my real portal, the exact problem that started this. Voice commands work in Egyptian Arabic, not just English. The hybrid split is a real design choice, not just a fallback: private by default, cloud only when it’s actually needed. You can trust it, because every deadline shows where it came from. What we learned Small on-device models are great for short, private tasks, but you have to design around their limits instead of hoping they’ll cope. Asking an AI to prove its answer (quote the source) removes most hallucinations. Students don’t need another place to upload files. They need something that works with where the files already are. What’s next for DueHunter Release it to students at my university on the real portal. Read university email too, for announcements and TA replies. Sync to Google Calendar and send reminders the night before. Sync across devices with Firestore. Merge duplicate deadlines that are announced in two places.