AI agents for biosensing
A measurement is not a decision. A lateral-flow strip, a paper microfluidic chip, or a wearable produces a number, and somebody still has to work out what it means and what to do next. I design grounded language-model systems to close that gap. In the framework, published as a Perspective, the system reads raw sensor output directly, checks it against a version-controlled knowledge base rather than against whatever the model happens to remember, and returns retest, refer, or report. A preprint carries the same idea into CRISPR diagnostics, with explicit bounds on what the agent may decide on its own.
The hard part is refusal. A system that guesses when the strip is smudged or the lighting is wrong is worse than no system at all, so quality control and human-in-the-loop oversight run across the whole path rather than sitting at the end of it. The target is the setting where there is no trained operator and no second opinion available.
Read Perspective, Frontiers in Sensors 2026 Preprint, chemRxiv 2026: AI agents for CRISPR diagnostics