Software Engineer, Model Capabilities

About the role

Notion has been at the cutting edge of AI since before ChatGPT launched. The job of the Model Capabilities team is to keep us there. We own the model layer of Notion AI: integrating frontier models as they ship, keeping inference reliable and economical at scale, and building new capabilities that other teams take advantage of. This role can be based in either San Francisco or New York City. We work from our offices on Mondays, Tuesdays and Thursdays (our Anchor Days) because we do our best thinking and building together in person. We’re looking for someone who’s excited to work alongside the team during those days.

What You'll Achieve

• Bring new frontier models into production quickly, making them available for our users and our engineers.
• Make inference reliable: better error categorization, self-healing retries, and cross-provider failover.
• Own observability for the model layer, driving down both time to detection and time to fix.
• Build new model-level capabilities and help product teams adopt them.
• Act as connective tissue across Notion's AI teams: find the gaps, unblock people, and make sure fixes land with the right owner.

Skills You'll Need To Bring

• Comfort with a moving frontier: The model landscape changes monthly. You keep up with it, form your own view of what's real, and stay pragmatic about what's worth adopting now.
• AI fluency as a practitioner: You use AI tools seriously in your own work and have opinions about where they help and where they don't.
• Production AI systems experience: You've built and operated systems on top of LLM APIs, not just prototyped with them. You understand token economics, streaming, tool calling, context limits, and how each of these fails in production.
• Measurement before conviction: You don't ship a prompt or model change on vibes. You define what "better" means, build or use the evaluation to test it, and are willing to be proven wrong by your own data.
• Cost and performance instincts: You treat latency and dollars as product qualities, not after