Inspiration

Small team projects involve more than completing individual tasks. Someone has to break down the requirements, assign responsibilities, pass results between people, and review the final work. Adding AI introduces another question: which parts should an agent handle, and where should a person step in?

We built KNOT to make those decisions visible. Our goal is to help small teams turn a project idea into a clear workflow that connects people, AI agents, and deliverables.

What it does

KNOT turns project requirements into visual AI workflows that users can edit, export, and run on their own computers.

Users can start with an idea, build a workflow from scratch, or upload a PDF, TXT, or Markdown requirements document. AI helps identify tasks and acceptance criteria, which users can review before generating a workflow.

On the canvas, users can:

  • Assign AI agents and human owners to tasks.
  • Choose different models and providers for different agents.
  • Run independent tasks in parallel.
  • Connect results from one task to another.
  • Add conditions and human-review checkpoints.
  • Define final Markdown, CSV, or JSON deliverables.

Once the workflow is ready, users can download a ZIP containing the workflow and a Python runner. It can execute independently of the editor, with model credentials configured separately.

How we built it

We built the interface with React, TypeScript, React Flow, and Vite. A Python backend handles document extraction, AI planning, workflow validation, and ZIP packaging. We use pypdf to extract text from uploaded PDFs.

The canvas and execution engine share the same JSON workflow format. This keeps the diagram connected to what actually runs.

Our model integrations allow agents to use different providers, including OpenAI and DeepSeek. The exported runner uses Python's standard library, and exported workflows contain credential references instead of API keys.

Challenges we ran into

One challenge was making every connection on the canvas have a clear meaning. We separated execution order from data access: a task starting after another task does not automatically mean it receives that task's results.

Another challenge was making human review part of execution. The runner needs to pause, save its progress, and resume after an explicit decision without repeating completed tasks.

Supporting multiple model providers also required handling different request formats, tool calls, and errors while preserving each agent's configuration.

Accomplishments that we're proud of

We connected requirements intake, editable planning, visual workflow design, and portable execution in one project.

We verified live workflows using OpenAI and DeepSeek together, including tool use and passing results between agents. We also tested exported workflows pausing for review and producing report and CSV files after approval.

Our automated checks cover the backend, workflow rules, and browser interactions, including document uploads, canvas editing, saving, and exporting.

What we learned

Building an AI workflow tool means making responsibilities, inputs, and review steps clear. A capable model still needs the right context and a well-defined task.

We also learned to test the complete journey—from editing a workflow to running the downloaded package—because a working interface alone does not prove that the workflow executes correctly.

What's next for KNOT

We want to add more reusable workflow templates, clearer execution progress, and better feedback when a task fails. Longer term, we would like to explore shared team workspaces and additional integrations while keeping workflows understandable and editable.

Built With

Share this project:

Updates

Submission history