Habit's

Learn
Class Schedule
Money
Routine
Ask
Agenda
App Icon

Inspiration

A student's week lives in too many places: assignments in Canvas, meetings in Outlook and Teams, deadlines buried in email, workouts in a notes app, event flyers on a hallway wall, money in your head. None of them talk to each other, so one change quietly breaks everything else. When a lab moves to Friday, nothing tells your gym plan, your shift or your budget. We wanted one app that connects those pieces and shows what a change actually affects before you commit to it.

What it does

Habits: a student starter pack is an iPhone assistant that pulls your classes, Canvas deadlines, calendar events, email (including Teams notices), habits and spending into one daily plan.

Ripple Mode: say "my lab moved to Friday evening" or ask "what if I go to the concert?". You see what it collides with, real free slots and what it costs against your weekly budget, then accept or reject the revised plan.
Campus Capture: photograph an event flyer. It reads the details, checks your calendar, your interests and that week's budget, and points out what the flyer leaves out, like the price.
Habits and classes: tick a habit and it slides away, cancels today's reminder and is remembered. Classes show on the day and the day before, for the whole term.
Money: log spending by voice, text or receipt photo. You get a weekly budget, upcoming bills, planned shift pay and an 8-week trend.
Ask: chat or talk to it. It answers from your agenda, email, budget and history, so "when did I submit my normalization lab?" gets an actual date. It reads replies aloud.
Memory Inspector: tap Why? on any suggestion to see the memories it used, then correct or forget them. The suggestion updates.
Spoken briefing: one tap reads today's plan out loud.

How we built it

The app is native SwiftUI. It talks to one FastAPI server on Vultr, behind Caddy HTTPS and token auth. On the server:

Gemini understands messy text, voice notes, flyer and receipt photos and email, and writes summaries and answers. It falls back across six models when one is busy or out of quota.
Tiger Data (Postgres + TimescaleDB) stores cached summaries and time-series hypertables for voice logs and spending. A time_bucket query powers the spending trend, and a history table answers "when did I…" questions.
Backboard holds long-term memory that the user can inspect and edit.
ElevenLabs voices the briefing and chat replies.
The Gmail API, with read-only scope, brings in email and Teams notifications.

Conflict detection in Ripple is plain, deterministic overlap code with a self-check. The model only parses the input and has to list its assumptions.

Challenges we ran into

Gemini's free-tier quota ran out mid-testing, and the SDK's silent retries made one request hang for over 4 minutes. We built a six-model fallback chain with a 25-second timeout per model.
Our school blocks direct Teams access, so Teams arrives the way students actually see it: as notification emails read through Gmail.
Keeping AI honest about dates and money. The model kept wanting to guess. We enforced "read, never guess" on the server: no amount means the app asks "How much was it?", and no date in the email means no date in the app.
iOS gotchas:
- a launch crash from two singletons initializing each other;
- the microphone staying on after dictation, which also silenced playback;
- code signing breaking because the project lived in an iCloud-synced Desktop folder;
- a Stats screen taking 19.5 ms per render, which we cut to 2.5 ms.
Backboard's free tier doesn't include its LLM, so we used it purely as a memory store and kept reasoning in Gemini.

Accomplishments that we're proud of

Ripple Mode works end to end on real data, and every link in a ripple traces back to a fact rather than a model's guess.
The memory is visible: correcting a memory in the Memory Inspector really changes the next recommendation in front of you.
Every sponsor service does a real job in a deployed, working system, not a demo stub.
Ask answers "when did I…" with real dates, and says "I don't have any record" instead of making one up.
It's a polished, native app running on a real phone: swipeable tabs, dark mode, reminders that work offline, and a bright new identity.

What we learned

Let models parse and deterministic code decide. Trust comes from showing assumptions and computing conflicts, not from asking an LLM to be careful.
Separate recorded, planned and hypothetical data. Money and schedules get confusing fast if you don't.
Free AI tiers need resilience by design: caching, fallbacks and timeouts mattered as much as prompts.
Small UX details decide whether people trust an assistant: a mic that stops by itself, a check that slides away, a button that says "Loading…".

What's next for Habits

Canvas API sync, so submissions are recorded automatically.
A home-screen widget for today's plan.
A side-by-side What-If view to compare two versions of your week.
An offline queue, so nothing gets lost without signal.
Smarter budgets from your real spending patterns.
Production: paid Gemini quota, verified Google sign-in and a TestFlight beta for classmates.