Omnicell AI

Inspiration

The idea came from a common problem in bioinformatics: many tools for single-cell visualization are slow and difficult to use. Moreover, as a bioinformatic scientist, sometimes you are doing standard actions to prepare the report about the data to share it with colleagues or other people, and usually these actions could be automated. Sometimes, colleagues can go back to you and ask some additional questions about the data, for which you usually should open code and load the data, but actually, this part could be also easelly automated with LLM. In this project, we tried to solve all these problems.

What it does

Our project helps users quickly explore their data, create visualizations, and generate reports for analysis. It can also connect to a local LLM, so users can ask questions about their data without sending sensitive datasets to the Internet.

How we built it

In the beginning, we used ChatGPT 5.6 Sol with a high-reasoning model to specify the best stack for rendering a dataset containing more than one million cells in a web browser. After the stack was specified, we split the large task into smaller ones to improve the performance of coding agents and, at the same time, remain fully involved in the project development. This was crucial for us because we wanted to always maintain a clear mental model of what our application was doing.

Each individual step followed this cycle, using Codex CLI in the beginning and later Claude Code after we ran out of tokens:

Planning → Implementation → Revision

Challenges we ran into

During the work, some features were not implemented efficiently or lacked user-friendliness. Sometimes, the first solution attempt was not optimized for performance or RAM usage and could cause memory overflow. In these cases, we had to create a new plan focused on improvements. After several cycles of such iterations, we steered the agent toward a good outcome.

Accomplishments that we're proud of

We are proud of building a real scientific application that can be used by bioinformaticians and biologists who do not have coding experience. One of the best parts of the application is that it is capable of working with huge datasets on local machines. This preserves security and prevents sensitive data from being sent to the cloud.

What we learned

While working on this project, we learned how to set up coding agents in an efficient and secure way using Docker Sandboxes, allowing us to use YOLO mode without fear of losing important data. We also learned how to use different models optimally depending on the task.