home page
ActionBrief turns rough meeting notes into a clear summary, decisions, open questions, and action items with owners and due dates

Inspiration

Meetings often produce useful ideas and decisions, but rough notes can make it hard to remember who agreed to do what. I built ActionBrief to make the follow-up clearer and easier.

What it does

ActionBrief takes pasted meeting notes and creates a concise summary, confirmed decisions, action items with owners and due dates, open questions, and the next meeting when it is mentioned. Users can download the brief as a Markdown file.

How we built it

I built the app with Python and Streamlit. When a user submits notes, ActionBrief sends them to an NVIDIA Nemotron model through the Nebius Token Factory API. The prompt asks the model for a structured response, which the app displays in separate sections. If an owner or due date is missing from the notes, the app marks it as unassigned instead of filling it in.

The deployed app reads its API key from Streamlit Secrets. It does not store meeting notes in a database.

Challenges we ran into

The hardest part was making the brief useful without letting the model turn guesses into facts. Meeting notes often mix confirmed decisions with ideas still being discussed, and may mention a task without a clear owner or deadline. I designed the prompt and output structure to separate decisions, action items, and open questions; it also tells the model to leave missing owners or dates blank and distinguish proposals from confirmed decisions.

I also had to make the model response usable by the app. The interface expects structured JSON, but model responses can include extra formatting, so I added handling to extract the JSON before displaying each section. Finally, I ran into a model ID that Token Factory could not find, so I had to select an available NVIDIA Nemotron model and keep the API key in Streamlit Secrets for deployment. These steps got the complete notes-to-brief flow working, though people should still review the output before sharing it.

Accomplishments that we're proud of

I built and deployed a working app that takes notes through a real Nebius Token Factory model call and returns a useful meeting brief. I’m proud that it separates decisions from action items and open questions, and that users can download the result.

What we learned

I learned that model IDs must match the models available to the Token Factory account, and that API keys should be stored in the hosting platform’s Secrets settings instead of the public code repository. I also learned to ask the model for a structured response so the app can present the results consistently.

What's next for ActionBrief

I’d like to add transcript file uploads, let users edit action items before downloading, and explore ways to send follow-ups to task-management tools. For now, users should review the generated brief before sharing it.