Joey Chakraborty

Building one AI agent for scientists and autonomous labs

Context
A growth-stage biotech company with an imaging platform, and a new direction for AI: bringing it to imaging.
Role
I conceived the agent and how it should work, won our VP of Engineering’s backing, led the hands-on build with a data science intern I hired, and presented our LLM plans to the board.
Impact
We took it from proof of concept to an in-development product that engineering now co-owns, built for a scientist or an upstream agent to drive.
Status
In development
Simplified concept view of the agent: chat on the left, with interactive parameters and approval on the right.

The situation

Imaging runs on intuition

Nothing was broken, which is why nobody asked for this. But surveying the company about AI showed me a chance to add precision to a workflow that runs heavily on expert intuition.

Before , a scientist sets imaging parameters from their knowledge of the sample, judges the images by eye, and decides whether to reimage or move on. No two scientists do it the same way, and the judgment takes experience to build.

Many imaging parameters can be inferred from the sample, such as cell size, dyes, and plate type, but the real test is whether the images are good enough for analysis. Our real-time analysis product, ClaireRT, answers exactly that question, so I saw that we could build an agent that checks its own work.

Early directions explored.

The work

A first bet we could finish

I chose imaging as a first project because its scope was more contained and better defined. Image analysis has a much larger scope and many possible methods, which means far greater complexity.

I convinced our VP of Engineering, my manager, that it was worth doing, and he let me hire a data science intern to build it with me.

Concept view of the evidence presented: reasoning, analysis results, and a review of the image before the driver approves or adjusts.

The work

Building trust: evidence for two drivers

I defined how the agent should earn trust before we built any of it. Whether a scientist or an upstream agent in an autonomous lab is driving, the result is judged from evidence:

  1. The agent’s reasoning for what it proposes, which includes how it interpreted the request and what it assumed.
  2. Real-time analysis of the images acquired with the agent’s imaging parameters.
  3. An optional review of the acquired images, by the scientist’s eye or by another agent.

Nothing runs until the driver approves it. If the evidence doesn’t hold up, the driver can take over and adjust the setup directly.

Who decides what: The model reasons and explains while our code calculates and analyzes in a predictable way.

The outcome

A new product in the works

The intern and I built a working proof of concept in under three months, and it was enough for our PM and executive team to see its potential and approve, on the spot, my request to extend the internship. Over the next few months, we iterated to refine the concept into a product; I decided where the model’s judgment ends and our code takes over, shaped the experience for the end-to-end workflow, and guided the intern through the build.

Engineering has now come on board to co-own it. We’re building it out as a solution that works as well with a scientist driving it as with an upstream agent in an autonomous, lab-in-the-loop workflow.