VerisFlow v0.4 Release: AI-Powered HSL Method Generation Achieves 92% Success Rate

1

Hi everyone,

I am excited to announce the release of VerisFlow v0.4, another step towards making AI a practical tool for generating HSL methods.

The journey with VerisFlow has been focused on steady, practical progress:

  • v0.2 proved that AI could generate HSL methods capable of running in simulation without errors.
  • v0.3 explored how to control the AI to produce these methods consistently, not just randomly.

Now, VerisFlow v0.4 verifies a crucial element for real-world use: setting precise pipetting parameters directly from a natural language prompt. This includes multi-channel volumes, tip types, liquid level detection, mixing cycles, and fixed/retract heights.

Key Features in VerisFlow v0.4:

  • Direct Worklist-to-Method Generation: Simply attach a worklist file, and VerisFlow automatically generates the corresponding HSL pipetting protocol.
  • Automated Verification: VerisFlow can now extract data from a trace log (.trc) and verify it against the original worklist to confirm accuracy.
  • Dramatically Enhanced Reliability: Our latest tests show that by optimizing prompt engineering, we have boosted the success rate of HSL method generation from 65% to 92% .

See it in Action: Check out this short video demo to see how it works: https://youtu.be/WMOElF0OVPc

The testing involved generating HSL methods from 100 randomly created worklists. Each generated method was then run in simulation, and the resulting trace logs were programmatically analyzed and compared back to the initial worklist.

A Note on Performance: Based on our 200 tests, the average time to generate a successful method is about 166 seconds . This requires an average of 1.3 to 1.7 API calls, as occasional failures may necessitate a retry.

For a deep dive into our methodology, data, and a full breakdown of the results, please see the resources below:

Feel free to try VerisFlow at www.verisflow.com.

I’m looking forward to your feedback and hearing about your experiences.

Best regards,
Xianghua

image
image1897×870 53 KB

6 Likes
2

Hi @Xianghua,
that’s very impressive work. I watched the Youtube videos and like your approach for simple tasks like pipetting from pos A to B or transports. However, I wonder if more complex routines can be created, like ones with loops, if/else structures or user defined error handling.

Another suggestion: Is there an option to create .med from the .hsl? As many method developers work with the method editor (more likely if they are no coders but scientists), that would be a plus.

one feature or tool I would really love to see: using your AI based tool to write HSL unit tests (in a respective unit testing framework that can e.g. be run in a git pipeline with HxRun control in simulation mode) for a given .hsl. I think that would close the loop as you can create the files with VerisFlow AND automatically ensure test coverage and to allow for creating modular HSL softwares.

cheers
Max

1 Like
3

Hi @max,

I really appreciate your valuable feedback.

The AI software currently handles loops and if/else statements. It does not yet support user-defined errors, as I had not included that use case originally. I plan to add the user-defined error handling per your suggestion.

Unfortunately, I haven’t resolved the difficulties regarding .med file generation yet. Actually, I have been working on an alternative solution. After uploading a protocol, the AI software generates a flowchart. Users can use this flowchart to adjust the workflow and parameters directly, instead of modifying the code manually.

The unit test you mentioned is a fantastic idea. I have developed a pipeline to verify the generated HSL methods. However, due to the lack of native APIs, this currently relies on GUI automation (interacting with HxRun by simulating mouse and keyboard inputs) and verifies the method by analyzing trace log files and comparing them against the input worklist. I’d love to develop a dedicated unit test framework.

Best regards, and Happy New Year!
Xianghua

1 Like