The Visionary Angle: OncoSight AI

Title

Inspiration

Biomedical researchers face a massive bottleneck when trying to leverage world-class datasets like the Imaging Data Commons (IDC) and the PanCancer Atlas. To extract meaningful insights, they typically must understand complex database schemas, write intricate SQL queries, and manually bridge the gap between medical imaging metadata and molecular/clinical variables. We wanted to eliminate this friction entirely. Our inspiration was to build a true AI data scientist—an autonomous, conversational agent that empowers researchers to explore complex oncology data and formulate radiogenomic hypotheses using plain English, all without writing a single line of code.

What it does

OncoSight AI is an autonomous biomedical data science agent that operates over the CRAFT semantic data platform. When a user asks a plain-English analytical question, the agent does not just guess an answer; it actively investigates. It discovers live connections to the IDC and PanCancer Atlas, scopes the underlying schemas, and uses Text2SQL capabilities to generate precise database queries. The agent then executes these queries, pages through the results, dynamically generates Plotly charts, and synthesizes the findings into a comprehensive markdown report. Crucially, it provides a fully auditable research trail by preserving the reasoning trace, the generated SQL, the raw data payloads, and the final visualizations for every investigation.

How we built it

We built the application using Python and a Streamlit frontend to provide an accessible user interface. The core intelligence is driven by Nebius LLMs (via Token Factory) or Anthropic models, orchestrated through a robust tool-use loop. We integrated the CRAFT Model Context Protocol (MCP) server to handle the heavy lifting of data interaction. The agent leverages specific CRAFT MCP tools—including search_schema, get_schema, sample_data, generate_sql, execute_query, get_result_page, and generate_plotly_chart. We designed the system to manage dataset-scoped Text2SQL, ensuring that queries against the IDC and PanCancer Atlas remain strictly separated until valid cross-dataset linkage is established.

Challenges we ran into

One of our primary challenges was preventing the LLM from "hallucinating" database structures or falsely claiming that disparate datasets could be easily joined. We had to strictly prompt the agent to perform schema discovery before making any analytical claims. Additionally, managing the live database-level scopes and child schemas (such as IDC_V17 and PANCANCER_ATLAS_FILTERED) required precise orchestration within the CRAFT MCP environment. Ensuring that the agent maintained context across multiple tool calls—from initial schema search to final chart generation—while keeping the IDC and PanCancer connections distinct was a complex architectural hurdle.

Accomplishments that we're proud of

We successfully delivered an end-to-end, working agentic workflow that interacts with live, massive-scale biomedical data. In our live smoke tests, the agent autonomously navigated the CRAFT connection, identified the correct schema, generated the appropriate SQL, and successfully retrieved records (such as identifying 142 imaging collections from the IDC) entirely from a natural language prompt. We are incredibly proud of the agent's auditability; by automatically saving the reasoning trace, raw SQL, and JSON data alongside the final report and charts, we built a tool that researchers can actually trust.

What we learned

We learned that when building AI agents for specialized domains like oncology, the semantic layer and schema discovery are just as critical as the underlying LLM. A reliable agent requires strict constraints, clear provenance, and transparent artifacts to be useful in a scientific context. We also realized that true cross-dataset analysis requires a careful separation of verified data links from exploratory hypotheses, highlighting the importance of robust tool orchestration over simply relying on model intelligence.

What's next for The Visionary Angle: OncoSight AI

Our immediate next step is to implement verified identifier mapping to enable robust, automated cross-dataset cohort building between imaging (IDC) and genomic/clinical (PanCancer) data. We plan to expand the agent's capabilities to include richer, cancer-specific analytical workflows, automated statistical significance testing, and the ability to export investigations directly into reproducible Jupyter notebooks. Ultimately, we envision adding collaborative workspaces and integrating broader biomedical datasets, transforming OncoSight AI into an indispensable, collaborative research partner for oncologists and data scientists worldwide.