totemBioinformatica

• 

  First screen after project creation

• 

  main project view, with agent on the right (like windsurf, cursor, etc.)

• 

  data preview

• 

  a generated analysis

• 

  The same analysis after user iteration

Inspiration

I worked as part of a 3 man team composed of a Clinical Doctor, myself (Data Scientist), and a Statistician. We would waste insane amounts of time coordinating across each other trying to create a single publication. It took us upwards of 2 years to complete our work, and this is after ChatGPT was already out there. Often frustrated by waiting on us to deliver results, the Clinical Doctor would try to use ChatGPT to do the analysis, but it simply couldn't cope with the scope of the study. On the other hand, I was able to use ChatGPT to write code (and execute it elsewhere) that could readily do what the Doctor wanted to do.

In a flash of inspiration, I realized that I was just an interpreter layer that enabled the Doctor to use ChatGPT to perform Data Science, and totemBioinformatica is an attempt to create an agent as that layer instead.

What it does

It takes the user's data and research goals, runs it across existing evidence based research pulled from the web, generates a study plan, then writes and executes code to conduct the study. It finally compiles the generated analyses and writes a publication grade report. The user is locked in to all sub-steps of the study plan, and can supervise or edit steps to ensure that the model defaults to the more relevant domain expertise of the user if necessary.

How we built it

totemBioinformatica was built using Next.js/React for the front-end and cloudflare workers for the backend. The UX is based on the years of experience I have working with domain experts and the knowledge of what is relevant to them and how they like to see it. Everything is on the front-end to allow for the data security concerns of users working with sensitive/protected data, and the LLM does not see the data itself, just a schema of the table columns.

Challenges we ran into

I learned typescript/javascript as I was building this and only recently (two-weeks ago) became aware of Next.js's routing and context system. I migrated to Next.js from a basic CRA framework and it took a lot of time, so the current implementation lags behind in some aspects to what I was able to accomplish before.

Accomplishments that we're proud of

I managed to build this myself and accomplished the basic proof of concept. As a use-case example, I can now perform survival-curve analysis on complex medical datasets in a matter of minutes where I remember working on them for days before.

What we learned

Over the course of building this, I learned how to effectively incorporate user feedback to iterate on my builds and expand the scope of my implementation in a sustainable way. I also became familiar with browser based stacks and how to use them. In the course of this hackathon, I learned how to incorporate perplexity's capabilities into my project and how to use them to enable real-time relevance of the analysis approach my agent takes. I also learned concepts unfamiliar to me about software engineering from the osmosis I experienced at the event.

What's next for totemBioinformatica

We are now going to work on polishing our features to the point that we can pitch this to customers as a complete product. I am still looking for a CTO, and will try to build an executable solution capable of using the terminal so that advanced species of research, such as genomics, can also be done using this platform.

Built With

• anthropic
• cloudflare
• grok
• messages
• next.js
• react
• typescript