Nova — AI Agent Platform for Data Science

The challenge

Data work is rarely contained in one interface. A useful assistant must understand the request, reach the right data, write and execute code, handle long-running compute, and return something a person can inspect — without hiding what happened.

Nova was created as an intelligent agent inside a data science platform. Its goal is to make that workflow feel coherent while preserving the flexibility that technical users expect.

What I worked on

• Shipped an MCP server and integrations for ChatGPT, Claude Code, and Cursor so data scientists can work from their preferred assistants and editors.
• Built a streaming chat experience that can generate multiple artifacts around a Jupyter notebook-centered workflow.
• Enabled users and teams to create, clone, customize, and share agents with system prompts, knowledge context, and reusable agent skills.
• Connected natural-language requests to high-performance computing resources for large, heterogeneous data analysis.
• Contributed across the web platform, CLI, AI platform, background workers, Python API package, and public MCP server.

System shape

The platform can be understood as four connected layers:

1. Conversation — stream intent, progress, code, and artifacts back to the user.
2. Agent — combine instructions, context, tools, and reusable skills.
3. Execution — run generated code against data and compute environments.
4. Ecosystem — expose the same capabilities through the product, CLI, API, MCP clients, and community-created agents.

The difficult work lives at the boundaries: keeping execution observable, carrying context between tools, and making a broad system feel predictable.

Current outcome

Nova is a working part of the platform ecosystem, supporting data interaction, artifact generation, agent customization, and access from external AI clients.

Content to add: architecture diagram, representative workflow screenshots, adoption or scale metrics, latency and reliability targets, and one detailed technical decision with its trade-offs.

What I learned

An agent becomes useful when its surrounding system is designed as carefully as the model call. Execution, feedback, permissions, and recoverability are product features — not implementation details.