Kenny Workman is a co-founder and chief technology officer of LatchBio, where he builds scientific software and benchmarks that test whether AI agents can produce defensible results from biological experiments. His central idea is that verifiable biological data analysis can give scientific agents what executable code gives software agents: a concrete way to measure whether their work succeeds.
Workman studied bioengineering and computer science at Berkeley and worked with Berkeley Lab, the Joint BioEnergy Institute, Asimov and Serotiny before leaving university and founding LatchBio with Alfredo Andere and Kyle Giffin. The company initially built browser-accessible computing infrastructure that helped biotechnology researchers store experimental data and run bioinformatics workflows without managing their own cloud systems.
That infrastructure eventually became the foundation for agents that analyze experimental datasets and dispatch computational work. In 2025, Workman developed latch-curate, a human-in-the-loop framework for single-cell data curation covering count-matrix construction, cell typing and metadata harmonization.
His independent study and writing span algebraic topology, category theory, statistics, immunobiology, developmental biology and base editing. He argues that engineers moving into biology must learn its concrete molecular vocabulary—including cell types, signaling pathways and immune proteins—because mathematical sophistication alone cannot replace scientific domain knowledge.
Making scientific AI measurable
- SpatialBench: Workman co-authored a benchmark for spatial-biology agents containing 146 tasks across five experimental technologies. Each task combines real biological data, a scientific objective and a deterministic grader, testing whether agents can extract meaningful results instead of merely answering textbook questions.
- SpatialBench Verified: Expert review of an expanded benchmark reduced 159 tasks to 115 reproducible evaluations. The process exposed ambiguous instructions, arbitrary quality-control thresholds and overly narrow grading tolerances—evidence that deterministic scoring is unreliable when multiple analytical methods yield scientifically valid answers.
- scBench: Workman co-authored 394 single-cell sequencing evaluations spanning six platforms, demonstrating that agent performance varies with experimental technology and available documentation.
- SpatialBench-Long: His long-horizon benchmark tests 24 extended investigations involving tumors, organoids, lineage tracing and aging. Tasks include reconstructing how a tumor seeded metastatic growth, requiring agents to connect experimental context, multiple assays and scientific interpretation. The strongest reported model-and-harness combinations succeeded in only 11.1 percent of runs.
Workman’s goal is reproducible biological agent evaluation grounded in deployed research workflows: identify where models fail, build rigorous tests around those failures and translate measurable improvements into more trustworthy scientific tools.
