Deadline Hive

Inspiration

Every student at COMSATS knows this feeling: a quiz or assignment gets posted on Teams, or Google Classroom, or the university Portal, or just dropped in a WhatsApp group — and it takes ten minutes, so you tell yourself you'll do it later. Then it's scattered across four different apps and you genuinely forget it exists until it's due. I'm a BS Computer Science student (FA24-BCS-063, section BCS-5B) and I've watched this happen to my whole class, repeatedly, on things that take almost no effort once you remember them.

The actual problem was never "I don't have time." It was "the deadline lives in a place I don't check." So instead of building another personal to-do app, I wanted something that turns the one thing every section already has — a WhatsApp group full of people who did see the notice — into a shared, AI-verified source of truth.

What it does

DeadlineHive is a crowdsourced, AI-parsed deadline tracker for a university class section. You join by name and email, and DeadlineHive suggests your section from your roll number (FA24-BCS-063 → BCS-5B) — matched against the real Fall 2026 CS department timetable, not guessed. You can join or drop individual courses too, which matters for a real case I have myself: repeating Data Structures through BSE-3B while my home section is BCS-5B.

Everything due across every section and course you're in lands on one searchable, filterable list — by section, course type, format, and done/to-do status — instead of hunting through separate pages. To add something, paste text or a screenshot from Teams, Classroom, the Portal, or WhatsApp. Gemini extracts the course, title, due date, and type/format tags, and checks it against what's already there — showing why it thinks something's a likely duplicate or a conflicting detail before you confirm. Nothing posts silently. Once confirmed, it's live for the whole section, your contribution count goes up, and anyone can flag it correct or wrong, or mark it done with the time it actually took.

How we built it

Planned entirely with Devpost's official Learn Skill Pack, driven through Claude Code: 1-start → 2-scope → 3-prd → 4-spec → 5-build → 6-ship, with every consequential product and technical decision made by me and recorded with its reasoning in scope.md, prd.md, and spec.md in the repo.

Stack: Next.js (App Router), TypeScript, a local SQLite file via better-sqlite3, and the Gemini API for extraction and duplicate detection — one structured call that receives the new item plus the section's recent items as context, and returns the extraction and a duplicate verdict with reasoning together. The full Fall 2026 CS timetable is seeded — 51 sections and 87 courses with real instructors, not just my own two sections — so section-suggestion and course search are tested against real data, not a toy subset. Identity is a lightweight name+email in localStorage, sent in request headers rather than URLs, with basic hardening added once real users were going to touch it: IP-keyed AI rate limiting, a prompt-injection guard on anything fed to Gemini, and security headers.

No pre-existing application code was reused — the skill pack provided the planning framework and process, not app code.

Challenges we ran into

I started with no API key and no fixed idea — just a frustration — and the skill pack's flipped-interaction planning is what turned that into a scoped, buildable product instead of a vague feature list. Technically, the free Gemini tier turned out to be noticeably variable: the model named in the docs was too slow or failing often enough that the app now tries a small fallback chain of models with a time cap on each, rather than assuming the happy path. Matching real roll numbers to real timetable sections was also messier than the simple "intake term + arithmetic" heuristic I started with — some intake cohorts needed matching against the timetable's own section headers instead, since straight arithmetic got some semesters wrong. And the whole thing was built against a genuinely tight clock — about 3.5 days from first idea to submission.

Accomplishments that we're proud of

Seeding the whole real Fall 2026 CS department timetable — 51 sections, 87 courses, real instructors — instead of just my own two sections, so the section-join and course-search flows are tested against real university data at real scale, not a cherry-picked demo case. Getting the duplicate-detection step to show its reasoning instead of silently guessing. And building resilience into the AI call itself (the model fallback chain) once free-tier reliability turned out to be a real product concern, not just a backend detail.

What we learned

That a free-tier AI model's reliability is itself a design constraint, not an implementation detail you can assume away — and that showing why an AI flagged something, rather than just a yes/no, is what makes an AI-assisted step feel trustworthy instead of arbitrary.

What's next for Deadline Hive

A section leaderboard, real university-email verification, push/email/WhatsApp reminders as deadlines approach, support for other COMSATS campuses and other universities, and eventually extending the same AI-extraction engine to the internships/jobs side of the original idea — a bigger, more scattered problem than class deadlines, and deliberately out of scope here.

Built With

• antigravity
• better-sqlite3
• claude-code
• gemini
• google-gemini-api
• next.js
• node.js
• react
• sqlite
• typescript