ChemSnap
Accelerating chemical imports with AI tools
- Client
SciSure
- Role
Designer & Developer
Summary
We worked with health and safety staff at 11 academic institutions to reduce the burden of chemical inventory upkeep. SciSure designed and developed an LLM-driven photo import feature. Hundreds of chemical container labels were used to supplement a general-purpose model that would determine manufacturer, quantity, product number, and chemical name from photos of chemical containers in the lab.
Outcomes
- Enabled for 131 customers since Nov 2024
- 557 chemical containers imported at 3 sites
Scope
Vitality Robotics developed a Flask API wrapper around ChatGPT that SciSure developers containerized and made production-ready. I developed front-end components to drop in to our legacy UI and consumed that API, delivering results to an existing Drupal form.
Process
For this project, we accelerated and expanded on our typical design process. Design at SciSure takes a cross-functional team of designer(s), product managers, and senior management. When I was hired, I was the sole designer on staff. I instituted weekly meetings of a core group to plan, spec, and iterate on potential designs. This team was dedicated to ChemSnap as our highest product prority.
There was market pressure to integrate LLMs into the platform, and we worked hand-in-hand with stakeholders from our leadership team throughout the process. This included inviting stakeholders to attend our research calls, including moderated remote testing sessions. We prefaced these with discovery interviews led by a designer (myself) and a product manager, aiming to meet with health & safety directors and principal investigators. We conducted 5 interviews with R1 research universities and a global pharmaceutical company to ensure that we would design for both subsets of our customer base — academic and commercial.
As an organization that considers itself a contributor to the safety and well-being of laboratory researchers, we have a responsibility to move carefully when considering the consequences that possible hallucinations and bad input could have when using AI for chemical inventory. ChemSnap AI fills but does not submit a form — intended to be confirmed by the user operating the camera.
Early Design
- Progressive disclosure avoids overwhelming the user and supports WCAG guidance for obvious next steps
- AI output is reviewed and manual conflict resolution avoids potential spam
- System stores input and output for future training efforts
We partnered with 11 early adopters to thoroughly vet the real-world performance of the feature and went thru multiple cycles of prototyping and iteration to arrive at the implementation we have today.
