← All projects

Clash of Plans Hackathons

Give a planning assistant something to point to

A student workspace, a saved task list, and answers that link back to those tasks.

An AI task and calendar agent.

GenAI Genesis 2025

GitHub ↗

The problem

A task list tells you what needs doing, while a calendar shows where it might fit. Clash of Plans, a GenAI Genesis 2025 project, brings tasks, visual day planning, focus sessions, and an assistant into one student workspace. This note follows its source-linked task-answering path.

How it works

  1. QuestionUser-owned tasks
  2. Keyword retrievalBounded context
  3. AnswerValidated task links
The task RAG path, separate from the original task-creation assistant.

  1. Load the signed-in user's tasks

    The service loads task records, then filters ownership again in retrieval. BM25, a keyword-ranking method, scores task titles, descriptions, and priorities against the question. Generic planning queries can fall back to priority order.

  2. Give the model bounded context

    Only a small set of matching task fields goes to the answer generator. This is retrieval-augmented generation, or RAG: find relevant records first, then ask the model to answer from those records.

  3. Validate the references

    The service checks the returned structure and task IDs. Cited IDs must exist in the retrieved set, and inline references must agree with the ID list. A response with missing or invalid references is rejected.

Decisions & tradeoffs

Keywords keep retrieval simple

Search needs no model call or vector database. It can miss a useful task when the question uses different words. Valid task references make an answer easier to inspect, but do not guarantee every sentence is true.

Be precise about persistence

Task records use Supabase. Dashboard day plans are browser-local per account and date, separate from Google Calendar. The README does not establish improved grades, wellbeing, or productivity.

Try it yourself

Read the retrieval and answer-service files linked below. The public landing page can also run from a fresh checkout with Node.js 20+ and npm.

cd client
npm ci
npm run dev

Open http://localhost:3000 for the landing page. Authenticated tasks and assistant answers need the separate backend, database, and provider setup. The landing page alone does not exercise those features.

Sources & scope

No local checkout was available. Public source was read directly. No measured user outcomes are claimed.