Dexter
An AI co-scientist: a one-sentence hypothesis → a costed, runnable experiment plan in 90 seconds.
A grad student with a Friday-afternoon idea needs a week of work before they even know if it’s worth running: read 30 papers, find a working protocol, source reagents with real catalog numbers and prices, build a defensible budget, and define what success means. Most ideas die there.
One API call returns a strictly-typed ExperimentPlan: a novelty check with linked references, surfaced hidden assumptions, a numbered protocol with citations, real reagents (Sigma / Thermo / Addgene, with catalog # + EUR price), equipment, a line-item budget, a phased timeline, quantitative success criteria, and likelihood-rated failure modes. Every field is schema-guaranteed via Zod + structured outputs, so the suppliers and catalog numbers are real, not invented.
Built for Hack-Nation 5 (Fulcrum Science track) with a live demo and public API. Compresses a week of pre-experiment work into ~90 seconds.
Schema-enforced outputs are the difference between an LLM that sounds right and one a PI would trust enough to order materials on Monday.