The next frontier of data
Delightful AI refines unpublished science and clinical data into frontier-quality RL environments at enormous scale, for the world's most important problems.
In 2021, the best AI coding tools could autocomplete a single line. Five years later, coding agents build and ship entire features on their own. Much of that leap happened because developers share millions of programs in public and every program can be checked by running it, giving models endless real problems to practice on and instant feedback when they get one wrong. Science and medicine have no such loop. There is no GitHub for experiments and no test suite for a diagnosis, which is why models that outscore PhDs on exams still struggle to plan an experiment or treat a patient whose case fits no guideline.
We believe science and healthcare are the most strategically important domains for the next phase of frontier AI. They are where the human and economic stakes are highest, where every major lab now has a dedicated mandate, and where models still fall furthest short of the best experts. They are also the hardest to train on, because of two constraints.
The first is fragmentation: the working record of science and medicine, from failed experiments and raw measurements to patient journeys and AI tool sessions, is scattered across thousands of organizations, and most of it has never been published or paid for. That record holds the failed attempts and judgment calls that papers and guidelines leave out, exactly what a model needs to learn how the work is done. Because it never reached the internet, no model has memorized it.
The second is verification: some outcomes are clean, like a reaction that works or a biopsy that confirms a diagnosis, but many arrive months later, are confounded by other variables, or never resolve. Therefore, whether an experiment was well designed or a treatment plan was right for a patient can only be judged by the few people best at that exact thing.
Solving both takes a company devoted entirely to the hardest problems in frontier data, which is why we are building Delightful AI. We source this data through our own infrastructure and refine it into RL environments with graders that frontier labs can trust.
Sourcing the most fragmented data in the world
This data sits with academic labs and research institutes, hospitals and health systems, biotech and pharma companies, contract research and diagnostic labs, and the AI tools scientists and clinicians use. No one has aggregated it, because no one has built the infrastructure or paid for it.
We are building both, with connectors that pull from lab notebooks, instruments, record systems, and product logs, and pipelines that handle consent, licensing, and de-identification so sensitive data never has to leave its owner. Partners share in the value their data creates, and we work with leading researchers and clinicians to collect new data built for training from the start.
Turning data into frontier environments
Frontier training needs environments of high quality and volume, delivered fast enough to keep pace with training runs. Today labs get two of the three: quality and volume arrive slowly, quality and speed come in small batches, and speed and volume sacrifice quality. We are building the lab and infrastructure to deliver all three, with expertise that deepens with every environment we build.
Our first job is deciding what is worth training on, since most data is too routine to teach a frontier model anything. The value sits in the informative minority: the experiment that failed for a reason nobody expected, the patient whose case didn't fit the guideline, the session where an AI tool confidently got it wrong.
Each research campaign or patient journey we select holds dozens of decisions that can each become an environment, and for sessions from deployed AI tools we rebuild what the user needed and the data and tools they had.
Scaling data quality is the bottleneck
The hardest part is the grader. Where reality returned a clean outcome, we freeze what was known at the time, ask the model to decide, and grade the decision against what happened. Where it didn't, leading researchers and clinicians define what good looks like, and we check their judgment against outcomes wherever possible.
We hold every layer to the same bar: the data we source, the environments built from it, and their graders all pass automated adversarial testing, agent-run quality audits, and expert review before reaching a lab, and our training runs show which data improves models.
AlphaFold was the payoff from half a century of the Protein Data Bank, where every structure met a common standard, and most of science and medicine never built anything like it. We intend to build it by combining what no one else does: proprietary access to fragmented data, the experts to grade it, and research infrastructure to produce environments and data to the highest standards at enormous scale.
Five problems we measure ourselves against
We are proposing five open problems that we intend to help solve. Each would change science or medicine if cracked, and each depends on exactly the kind of data and grading we are building.
- Catch toxicity before the first human dose. Preclinical tests miss up to half of the drugs that later injure the liver, and safety failures are among the costliest ways a drug program can end. A model that flags those drugs using only what was known before anyone was dosed would protect patients and spare years of wasted development.
- Predict how a cell responds to a perturbation it has never seen. In Arc Institute's 2025 Virtual Cell Challenge, models from more than 1,200 teams still didn't consistently beat naive baselines. A model that predicts what happens when a gene is switched off or a drug is added, in a cell type it has never seen, would let biologists run far more experiments in silico than any lab could run at the bench.
- Find antibiotics that can get into gram-negative bacteria. Drug-resistant infections kill more than a million people a year, and no new antibiotic class for gram-negative bacteria has reached patients in decades, largely because few molecules can cross their outer membrane. Predicting which new molecules will get in and kill would reopen one of medicine's most urgent pipelines.
- Make ammonia without extreme heat and pressure. Fertilizer from the Haber-Bosch process feeds roughly half the world's population, but the process runs at 400 to 600°C and 100 to 200 atmospheres and consumes about 1 to 2 percent of global energy. Bacteria perform the same reaction at room temperature, so a catalyst that matches them is possible in principle. Predicting which candidates will work, using the failed attempts sitting in lab records, could cut one of the largest energy demands in industry.
- See sepsis hours before it sets in. Sepsis is linked to about one in five deaths worldwide, and a widely used hospital warning model missed two-thirds of cases when it was independently tested. A model that warns clinicians hours earlier, with few enough false alarms that they act on it, would save lives in every hospital that adopts it.
The future of science and healthcare AI
We chose science and healthcare because their frontier is unbounded. Long after models master coding and knowledge work, every discovery will raise new questions and every patient will bring a new case, and we will be there to keep frontier models learning.
If you hold research, clinical, or deployment data, lead your field, or want to help build this, let's chat: harshel@delightfullabs.ai