THALAMUS
An AI research assistant that finds proper drug candidates for pharmaceutical company
About the project
I’ll show a small experiment connecting a biological model to an AI agent.
The question is simple: can we give an agent numbers describing cells, let it use a calculator, and check whether its answer is correct?
We will watch the actual requests and results as they happen.
Built with
AI tools
Prize Categories Entered
- Submitted

Best Use of Gemini API
Gemini (gemini-3.5-flash-lite, structured output) is the agent's reasoning engine: it turns a question into a validated analysis plan, decides what the literature search is still missing across up to four search rounds, and writes conclusions in which every claim cites an observation or paper id. We also ran a controlled experiment on Gemini showing that giving it a trust report next to each number raised noise-aware answers from 2 of 12 to 11 of 12. Responses are cached and paced to stay within the free tier.
- Submitted
Best Demo Video Using HyperFrames
HyperFrames turns code into editable video, so we used it to explain our data, agent workflow, and results in one short demo.
- Event prizeSubmitted
Best Overall Hack
Its core idea is simple: never hand an AI a number without telling it how much to trust that number. Every result the agent sees comes with a "trust report": how large the natural noise is, how many times the experiment was repeated, and whether anything looks wrong with the data. We tested this on the same measurement, which was actually just noise. Given only the number, the AI called it a real effect in 10 of 12 tries. Given the trust report too, it correctly called it noise in 11 of 12. The agent is also honest about what it finds: for one cancer it found a drug that affects only the cells with the target mutation, and for the other it reported that no drug was selective, instead of making one up.
- Event prizeSubmitted
Cleanest Use of a Software Factory
THALAMUS demonstrates our approach to a clean software factory: define the expected behavior precisely, build components with clear responsibilities, and make the resulting behavior verifiable through tests and execution evidence. We applied this approach to a scientific agent that connects real Tahoe cell-state data, Gemini model calls, numerical tools, and an interactive research interface. We started with explicit contracts for the experiment. Each representative cell state must contain exactly 2,560 finite numerical coordinates. Candidate identifiers must be valid and unique. The agent can call only the permitted scoring function with the identifier of the current experiment. Its final response must contain an allowed candidate and an explanation in the expected structure. These contracts give the implementation concrete acceptance criteria and make malformed inputs or responses visible. The code separates data validation, numerical computation, model communication, experiment or