Inspiration

Important commitments are often made during everyday conversations, but they are easy to forget or lose track of. A simple sentence like “I'll send the quotation by Friday” can contain a real responsibility, deadline, and expected action.

We wanted to build a system that can understand these spoken commitments and turn them into concrete actions instead of leaving them buried inside conversations.

What it does

CommitmentOS turns spoken commitments into actions.

It analyzes conversations, identifies genuine commitments, and extracts important details such as:

What needs to be done
Who is responsible
Who the commitment is for
When it needs to be completed
Relevant context or dependencies

The extracted commitments can then be organized into actionable tasks, calendar events, or other follow-up actions.

How we built it

CommitmentOS is being built as an AI-powered application using Next.js and TypeScript for the application layer.

The planned processing pipeline is:

Conversation → Speech-to-Text → NVIDIA Nemotron via Nebius → Commitment Detection → Structured Commitment → Action

The AI layer uses NVIDIA Nemotron through Nebius Token Factory to reason about conversations and identify meaningful commitments rather than simply matching keywords.

Challenges we ran into

The biggest challenge is making the system understand the difference between a commitment and an ordinary statement.

Real conversations are messy. People use vague deadlines, change their minds, refer to previous conversations, or make conditional statements.

We also have to handle the technical complexity of connecting conversation audio, speech-to-text, AI reasoning, structured outputs, and external actions into one reliable workflow.

Accomplishments that we're proud of

Designed a workflow that connects conversations directly to actionable commitments.
Chose a structured representation for commitments instead of treating them as plain text.
Integrated the project architecture around NVIDIA Nemotron and Nebius.
Built the initial CommitmentOS application foundation with Next.js and TypeScript.
Focused on making AI output useful in the real world rather than just generating a conversation summary.

What we learned

We learned that building an AI application is more than connecting an LLM to a prompt.

The difficult part is designing the complete system around the model—how information enters the system, how the model reasons about it, how its output is structured, and how that output becomes a reliable action.

We also learned the importance of handling uncertainty instead of blindly assuming that every sentence represents a confirmed commitment.

What's next for CommitmentOS

Our next goal is to build the complete end-to-end workflow:

Live conversation → Speech-to-Text → Nemotron → Commitment Engine → Action

We want CommitmentOS to eventually understand commitment dependencies, handle ambiguous deadlines, maintain commitment history, and proactively identify commitments that may be at risk.

The long-term vision is to make CommitmentOS a reliable commitment-management layer for conversations, turning things people say into things that actually get done. 🚀