startItAp-orchestrator-hackathon2025-3

Inspiration

I often struggle with the same paradox: I have lots of ideas for side projects, but I lose momentum when I have to manually break them down into phases, tasks, tech stack decisions, and realistic roadmaps.

This project started as an experiment:

What if I could treat an AI model like a tiny “project orchestration team” that helps me go from messy idea → clear brief → concrete plan → executable tasks, all inside one workspace?

I also wanted something I would actually use for my own projects, not just a demo.

What it does

This app is a small AI-assisted project orchestrator:

You start with a short idea (e.g. “simple AI planner for dev projects”) and optional constraints (team size, timeframe, dev skills).

An AI brief agent turns that into a structured project brief (problem, solution, goals, technical context).

A planning + architecture agent generates:

project summary

phases with tasks and rough estimates

suggested architecture (style, modules, data flow)

a concrete, minimal tech stack (frontend, backend, DB, infra, testing/tooling)

The backend persists phases & tasks into a relational DB and exposes them via an API.

In the UI you get:

Dashboard – project snapshot, phases/tasks counts, workload breakdown

Tasks board – Kanban-like view, where each task can be opened and edited (title, description, priority, phase, status)

AI tab – buttons to:

re-run the planning agent (read-only preview)

apply the AI plan and overwrite phases & tasks in the DB

ask AI for project risks

generate a tech spec (Markdown)

generate full project documentation as a LaTeX .tex file

The LaTeX agent returns a full article document with sections like Introduction, Architecture, API surface, Data model, Risks & mitigation, and Roadmap – ready to drop into a repo or internal docs.

How we built it

Frontend

React + TypeScript

TanStack Query for server state

Tailwind CSS for styling

A small tabbed project layout (Dashboard / Tasks / AI)

Backend

Node.js + Express

Prisma ORM

SQLite for persistence (simple dev-friendly DB, easy to swap later)

OpenAI API (responses API) for all agents

Challenges we ran into

Reliable JSON from the model Getting the model to always return strict JSON with a fixed shape (no comments, no extra keys, no markdown) required careful prompt design and validation. I had to handle parsing errors gracefully and surface them in the UI.

Safe “apply” of AI plans Overwriting phases & tasks is destructive. I ended up implementing:

a read-only replan preview first,

then a separate /projects/:id/apply-plan endpoint that:

deletes existing phases & tasks for the project in a transaction,

recreates them from the AI plan.

Balancing simplicity vs. power It’s tempting to add drag-and-drop Kanban, multiple plan versions, advanced permissions, etc. I had to keep the scope small and focus on the core loop:

idea → brief → plan → stored phases & tasks → edits → replan/apply.