case study · AI · Scheduling · Gemini Hackathon

24Flow

An AI timetable that plans all 24 hours of your day.

Role
Solo: design, frontend, backend, AI integration
Timeline
Hackathon sprint, 2026
Status
Live
Links
live ↗code ↗
Stack
SvelteKitNode / ExpressPostgreSQLPrismaGeminiGSAP

Type your tasks in plain language and 24Flow turns them into hourly blocks, fills the gaps, balances focus and breaks, and resolves overlaps by priority.

01

The problem

Calendars assume you already know when you’ll do things. Most people know what they need to do, not when, and the moment two things collide the whole plan falls apart.

Who it’s for

Students and busy people who plan in to-do lists, not calendars.

02

Goals & constraints

Goals

  • Go from a sentence (“gym, 2 hours of revision, call mum”) to a full day plan.
  • Plan the whole 24 hours, sleep and breaks included.
  • Detect clashes and fix them in a way that respects what matters most.

Constraints

  • Hackathon timeline
  • AI output has to land as valid, non-overlapping time blocks every time

03

Approach

  1. 1.

    Natural language in, time blocks out

    A language service parses free-text tasks into items with duration and priority; an auto-fill service places them into free slots around fixed events.

  2. 2.

    Conflicts as first-class objects

    Overlaps are detected as clusters, stored as a Conflict with a readable ID and AI analysis, and the user gets a proposal they can accept.

  3. 3.

    Context beyond tasks

    Mood, sleep and behaviour logs are in the data model, so the planner can protect sleep and place focus work where it’s realistic.

04

Architecture

A cron service handles recurring events and reminders; an OCR service imports timetables from images.

  1. 01SvelteKit UI → Express API (JWT / Google auth)
  2. 02languageService → structured tasks
  3. 03autoFillService → places tasks in free slots
  4. 04conflictService → overlapping clusters → Gemini proposal
  5. 05overlapFixerService → applies accepted proposal → Postgres via Prisma

05

Key decisions

Priority-aware resolution

High-priority items stay fixed and low-priority ones move. A client presentation stays at 10:00; the coffee break moves.

Model fallbacks via config

The Gemini model and an ordered fallback list are environment variables, so swapping models never means editing features.

↳

DSA in the code

Real algorithms, solving real product problems.

Interval merging

conflictService

Merges overlapping busy intervals so the model sees clean free slots.

Sort + sweep for overlap clusters

conflictService

Events are sorted by start time and swept to group overlaps into conflict clusters.

!

Challenges

Trusting the model with time

LLMs are bad at arithmetic. The model proposes; deterministic code validates every block before anything is written.

✓

Outcome & lessons

  • Submitted to the Gemini hackathon with a live deployment and demo video.

What I learned

  • Let the model reason, let code enforce.
  • Interval problems show up everywhere once you look.

What’s next

  • Calendar sync
  • Weekly reflection from mood and behaviour logs
next projectImageFuse→