The YC startup says its models cut formulation work from years to weeks, and it is building a lab to generate proprietary training data.
By Ryan Merket · Published
Primary source: Y Combinator
Why it matters #
Rasyn is betting that proprietary lab data, rather than another chat interface, will create the defensible advantage in AI chemistry. The strategy can support larger contracts, but it also adds the cost and validation burden of physical science.
Ansh Tiwari, Ayush Chauhan and Daood Hashmi are expanding Rasyn from an AI tool that runs chemistry software into a company that designs, synthesizes and patents new chemical formulations.
The three-person San Francisco startup, founded in 2025 and part of Y Combinator's Summer 2026 batch, says it has already produced three previously nonexistent formulations. Rasyn identifies one as a PFAS-free coolant for data centers and says all three have been synthesized and patented. Those assertions come from Rasyn's YC profile; the public claim does not include patent numbers, formulation specifications or independent performance data.
The founders bring an unusually young technical team to the problem. Tiwari is a Caltech physics student who worked at NASA's Jet Propulsion Laboratory and helped develop sensor systems to reduce the energy consumed by laboratory fume hoods. Hashmi studies chemical engineering at Caltech, where his 2025 summer research examined palladium-catalyzed reactions in the Stoltz group. Rasyn says he has also worked on diffusion models for reaction mechanisms and synthetic pathways in the Arnold and Marcus labs.
From agent to formulation engine
Rasyn's current pitch is broader than the product it introduced on August 4th. In its initial YC launch, the founders described Marigold as an AI agent that accepts a chemistry task in plain English, searches the literature, reads laboratory files, selects software and runs calculations.
That version attacked the clerical work around computational chemistry. A researcher might otherwise move structures and results among separate programs for retrosynthesis, docking, spectroscopy and property prediction. Rasyn said Marigold could connect those steps, check intermediate results and stop when a chemical assumption failed.
Marigold Science remains a commercial product. Rasyn's site lists 54 chemistry and scientific-computing tools that the platform can run, including AlphaFold2, AutoDock Vina, RDKit, OpenMM and PySCF. Customers can also self-host Marigold so instrument data and unpublished results remain on their own hardware.
The pricing page starts at $89 a month for an individual researcher. A laboratory plan costs $249 per seat each month with a three-seat minimum, while institution plans start at $2,000 a month and support private model endpoints and self-hosting.
The updated strategy puts Rasyn further inside the scientific process. Rasyn now says it uses its own models to predict how mixtures will behave, ranks candidate formulations and validates selected candidates experimentally. It is also building a laboratory to produce the experimental data used to train later models.
That is a materially harder business than selling an agent subscription. It also gives Rasyn a route to data that competitors cannot collect by scraping papers or connecting existing open-source tools. Each physical experiment can become training data for the next prediction cycle, provided Rasyn can run enough experiments and measure them consistently.
The 100,000x claim needs a baseline
Rasyn headlines its YC profile with a claim that it can make chemicals "100,000x faster." The operational comparison underneath is narrower: Rasyn says conventional formulation development takes three to five years of laboratory trial and error, while its models can reduce the design process to weeks.
The 100,000x figure has no disclosed unit, baseline or test protocol. It could describe the speed of a model prediction against a laboratory experiment rather than the elapsed time required to deliver a commercially usable chemical. The years-to-weeks comparison is the more relevant measure for manufacturers, though Rasyn has yet to publish customer results establishing that cycle time across multiple projects.
Rasyn says it is already designing formulations for large manufacturers. Its public materials do not identify those customers or specify whether the work is paid, in testing or headed for production.
The company's published technical results offer evidence that the founders are building chemistry-specific models, while stopping short of validating the broader formulation claim. One company-authored ChromPeakNet study reports an F1 score of 0.889 on real chromatography data, compared with 0.476 for MZmine. The real-data benchmark contained 32 labeled peaks, a useful test with a limited denominator. Rasyn also publishes results for retrosynthesis, reaction-condition prediction and nuclear magnetic resonance shift prediction on its research page.
The laboratory becomes the moat
Rasyn is aiming first at formulations used in products such as data-center coolant, thermal paste, paint and adhesives. These are mixtures of multiple ingredients, creating a large search space in which changing one component can alter viscosity, heat transfer, stability, cost or safety.
The PFAS-free coolant project also places Rasyn in a market facing regulatory pressure. The Environmental Protection Agency expanded federal reporting requirements again in 2026, bringing the number of PFAS substances tracked through the Toxics Release Inventory to 206. A coolant that delivers the required thermal performance without persistent fluorinated chemicals would give manufacturers a practical reason to test an unfamiliar formulation.
Rasyn now has to demonstrate that its predictions survive synthesis, testing, scale-up and manufacturing. Success would let the founders sell higher-value formulation work while using Marigold as the software layer underneath their own laboratory. Failure would leave Rasyn with a capable chemistry agent and a headline speed claim that moved much faster than the chemicals.