A two-person YC startup is building agents that read the literature, run the experiments on borrowed GPUs, and hand back a paper draft. The wager underneath it all: the lab that captures the most research-process data wins the model.
Ask a working scientist how they spend a Tuesday and the answer rarely involves discovery. It involves tabs. A literature manager in one window, a Jupyter notebook in another, a GPU queue somewhere in the cloud, a half-finished draft in a fourth. Each tool knows its own slice and none of them talk. Every time the researcher switches, they rebuild the same context by hand: what they read, what they tried, why they chose this method, what broke. Synthetic Sciences, a company in Y Combinator's Winter 2026 batch, is built on a simple observation about that Tuesday - the bottleneck is not intelligence, it is fragmentation.
The company's pitch fits on a bumper sticker its founders actually use: "Cursor for Scientific Research." Point the product at a research question, a code repository, or a dataset, and its AI co-scientists take over the loop. They search and synthesize the literature, generate hypotheses tied to prior work, design experiment plans, write and run the code in containerized environments, launch jobs on rented GPUs, watch the results, and hand back a draft in LaTeX with figures attached. The human stays in the chair. The busywork moves off it.
Over the past couple of years, AI has posted visible wins in mathematics. That is not an accident, the founders argue - math is easy to score, so progress is easy to show. The rest of science is messier. Computational biology, proteomics, and machine-learning research do not reduce to a clean benchmark as neatly, and the tooling has lagged behind. Synthetic Sciences decided to plant itself in that harder, quieter territory rather than chase the domains that already photograph well.
The founding line the company keeps returning to is blunt about the constraint: you cannot parallelize yourself. A single research question can splinter into months of work scattered across tools, tabs, and time zones, and every step demands the same person's attention. The comparison the team draws is to software engineering, which got dramatically faster over roughly 24 months once agents could hold context across a codebase. Do the same for research, the argument goes, and the output is not faster typing - it is more shots on goal.
"You can't parallelize yourself. If we could accelerate scientific research the way we've accelerated software engineering in the last 24 months, we are guaranteed to see breakthroughs that massively benefit humanity." Synthetic Sciences, YC launch
The company is two people. Aayam Bansal, the co-founder and CEO, has a resume that reads like it skipped a decade. At 17 he patented - in the US, Singapore, and India - and sold an AI orthopaedic sock, and during the pandemic he built helpline infrastructure for the Indian government that served more than 50,000 people a day. His co-founder, Ishaan Gangwani, arrived from the competitive end of the field: USACO Platinum, which puts him in a rounding error of the world's competitive programmers, and an Honorable Mention at the 2025 International Olympiad in Artificial Intelligence.
They did not meet at a demo day. They met doing machine-learning research across the National University of Singapore, Carnegie Mellon, and MIT's CSAIL, and they published together at workshops for venues like NeurIPS, ICML, and AAAI before they ever incorporated. That matters for a product aimed at researchers: the people building the co-scientist have sat through the marathon it is trying to shorten. The company was known briefly as InkVell before taking the name Synthetic Sciences.
Public sources credit the founders with publications at NeurIPS, ICML, ICLR, AAAI, and CVPR workshops. Workshop and main-track acceptances are different animals; the distinction is worth keeping in mind when a two-person team cites a wall of A-list venues.
The flagship is the hosted platform, referred to internally as synsci. It runs in modes tuned to how research actually splits up - a Research mode for ML workflows, a Biology mode for lab-adjacent work, a Write mode for turning results into academic prose, and a Flywheel mode that is really a business strategy wearing a feature's clothes. Underneath, the system dispatches experiments across a spread of GPU providers, keeps persistent sandboxes so a long job can survive an interruption, and treats citations as a queryable graph rather than a pile of vector-search hits.
In July 2026 the company did something less common for a seed-stage startup: it open-sourced a large piece of the stack. OpenScience is a browser-based research workbench released under Apache 2.0, model-agnostic by design, so a lab can point it at Claude, GPT, Gemini, or a local fine-tune and switch per request. It ships with more than 250 editable skills and direct hooks into roughly 30 scientific databases - UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, and the rest. The framing is pointed: scientific AI tooling, the company says, should not be owned by one vendor. Atlas is the optional managed layer on top - curated frontier models, a persistent research graph, and hosted compute for teams that would rather not run their own.
"Scientific AI tooling should not be owned by one vendor." Synthetic Sciences, on releasing OpenScience
Flywheel Mode is where the company's real thesis shows. Everyone can rent the same models and buy the same tokens; what a lab cannot buy is another lab's corrections, accept-and-reject decisions, traces, and outcomes. Flywheel turns that exhaust into training data and handles the unglamorous plumbing - evals, deployment, compute across 20-plus providers - so a team can fine-tune models it actually owns instead of renting intelligence by the token. It is a switching-cost engine dressed as a convenience.
That 92% is the number the company leads with: on BixBench Verified, a computational-biology benchmark, its biology mode reports state-of-the-art performance. Most of the platform's testing so far has been on ML research, with biology the fastest-expanding front. A benchmark is a benchmark, not a Nature paper, but it is a concrete claim from a team that could otherwise be dismissed as two people with a landing page.
Synthetic Sciences is not alone in the water. Google has demonstrated an "AI co-scientist," Anthropic has moved toward science-focused tooling, and startups like FutureHouse, Elicit, and Consensus each own a slice of the pipeline. The incumbents of research operations - Weights & Biases, Jupyter - own the parts they own and little else. The company's argument for daylight is the loop itself: rivals tend to nail one stage, while Synthetic Sciences is betting that carrying context unbroken from literature into code into writing is the feature nobody else has bothered to finish.
| Capability | Point tools | Synthetic Sciences |
|---|---|---|
| Literature synthesis | Yes | Yes |
| Runs code & GPU jobs | Rarely | Yes |
| Context across the loop | Siloed | Shared |
| Own / fine-tune models | No | Flywheel |
| Open-source option | Varies | OpenScience |
Getting in is cheap by design. New users get $5 in credits at sign-up, and credits abstract the underlying GPU and model costs so a researcher does not have to think in tokens. From there the motion is bottom-up: individual researchers first, then teams, then university and enterprise contracts, with reported monthly subscription tiers and custom enterprise deals for the labs that need priority compute. There is a second, quieter revenue idea in the mix - selling reinforcement-learning environments and process-oriented research data to AI labs that want to post-train models on how research actually gets done.
The company raised a $1.4M pre-seed and joined YC's Winter 2026 batch, where its group partner is Gustaf Alstromer. Two people, a pre-seed, and a claim on one of the largest markets imaginable - all of human research - is either the setup for a punchline or the setup for a compounding advantage. The honest read is that the technical pieces here are not exotic; a capable team could assemble a rough version in a couple of quarters. What is hard to copy is time. If a lab's experiment history and fine-tuned models embed into the Flywheel, the cost of leaving climbs quietly, month after month, and that is the bet the whole company rests on.
"Nobody else can buy your users, your logs, and your feedback loops." Synthetic Sciences, on Flywheel Mode
The interesting question is not whether agents can draft a paper - they can, unevenly. It is whether Synthetic Sciences can turn everyday research use into a data advantage faster than a better-funded team can out-distribute it. Big platforms have reach the startup cannot match; the startup has focus and an open-source flag planted in front of the vendors. Ten years out, most of the specific tools in this story will have been renamed or absorbed. The durable question underneath - who owns the data that teaches a model to do science with taste - is the one worth tracking, and it is the one this company decided to build its life around.