COMPANY Health Tech · AI · Life Sciences
Frekil Wants Drug Evidence to Move at the Speed of a Chat Message
The Y Combinator-backed startup automates real-world evidence for pharma, compressing a process that usually runs 6 to 18 months. Its one hard rule: patient data never enters the model.
A drug gets approved. It goes into millions of bodies that never showed up in the clinical trial - people with three other prescriptions, a heart condition, a different metabolism. Somewhere in the electronic health records and insurance claims that pile up afterward, the real story of that drug is being written. The problem is that reading it takes a long time. Frekil, a two-person company from Y Combinator's Spring 2025 batch, is betting the wait can shrink from months to minutes.
The category is called real-world evidence, or RWE, and it is one of those unglamorous corners of an industry where enormous amounts of money and time get spent out of public view. When a pharmaceutical company wants to know whether a drug is causing a rare side effect, whether it works as well in older patients, or whether it deserves an expanded label, someone has to build a study out of messy, fragmented patient data. That work is slow, expensive, and mostly invisible to the people whose health depends on it.
Clinical trials answer a narrow question under controlled conditions: does the drug work in a carefully chosen group of patients. Real-world evidence answers the messier one that follows: what happens when the drug meets the actual population, with all its variety. Regulators increasingly want both, and the second kind of evidence gets harder to produce precisely because the data was never designed to answer research questions in the first place. It was designed to bill an insurer or chart a visit.
Frekil's pitch is that most of that slowness is not science. It is plumbing.
By the company's own accounting, a single real-world evidence study can cost between $100,000 and $1 million and take 6 to 18 months, and the majority of that effort goes into cleaning, joining, and reshaping data before anyone runs an analysis. Frekil's name is an argument in acronym form: Fast Real-world Evidence & Knowledge Insight Loop. The company wants to own the boring middle of that sentence, the part between raw records and a finished, defensible study.
What it doesAn automated biostatistician
In practical terms, Frekil connects to the data a life-sciences team already has - EHR systems, claims databases, registries, and proprietary datasets - and then lets researchers run a study by talking to it. You describe a hypothesis. The platform helps clarify the question, assembles a patient cohort, maps out the causal assumptions, drafts a statistical analysis plan, extracts the data, runs the analysis, and writes up the report. The company describes the experience simply: chat with your data, refine each step, and let the agents improve with every study.
Under that conversational surface sits a defined assembly line. Frekil breaks a study into eight stages, each handled by its own agent and each open to human review before the next one runs.
The design decision worth pausing on is the fourth box. A Causal DAG - a directed acyclic graph - is how epidemiologists diagram which factors influence which outcomes, so that a study does not confuse correlation with cause. Putting that step in the middle of an AI product is a signal about the audience. Frekil is not trying to hand a shortcut to people who want a quick chart. It is trying to speak the language of biostatisticians who will be asked, later, to defend every choice in front of a regulator.
The hard ruleNo clinical data touches AI
Every AI company in healthcare says some version of "we don't train on your data." Frekil states it as architecture rather than a promise. Its founding principle, printed plainly on the site, is that no clinical data ever enters an AI model. The agents orchestrate the work - deciding what to do and generating the code to do it - while patient-level records stay isolated from the model itself.
That separation is not just a privacy talking point. It shapes what the product outputs. Rather than returning an answer you have to trust, Frekil generates reproducible code - R, Python, or SAS - so a customer's own statistician can read it, run it, and audit it. In a field where a study might one day support a label change or a safety decision, "here is the code we ran" is a stronger position than "the model says so."
Who it's forThe buyers behind the science
Frekil's customers are the teams that live or die by this evidence: pharmaceutical and biotech companies, contract research organizations, and the health-economics and market-access groups that argue a drug's value to payers. Add the post-market safety teams watching for signals after launch, and the academic epidemiologists who run these studies for a living. The use cases the company lists read like a tour of where RWE money goes - HEOR and market access, post-market safety, label expansion, competitive intelligence, trial feasibility, and external control arms, where real-world patients stand in for a placebo group.
The fieldWhere Frekil fits
Real-world evidence is not an empty market. Established names like Aetion, TriNetX, Flatiron Health, Komodo Health, and IQVIA have spent years building data networks and analytics for exactly this audience, and plenty of studies still run the old way, inside CROs and in-house biostatistics teams doing the work by hand. Frekil is not arriving with a bigger dataset. Its wager is different: that the workflow itself - the plumbing, the drafting, the reformatting - is the part ripe for automation, and that an agentic, conversational layer can sit on top of the data a customer already owns rather than trying to out-collect the incumbents.
That is a narrower and, in some ways, humbler place to stand. Frekil connects to Databricks, Snowflake, AWS, GCP, and Azure and runs where the customer's data already lives. The product is the speed and the rigor, not the raw information. It is a distinction that matters for trust: a customer never has to hand its most sensitive asset - patient records - to a startup in exchange for a study.
The bet also fits the moment. General-purpose AI has made the drafting steps of knowledge work - summarizing literature, writing a plan, generating code - dramatically cheaper. What it has not solved is the surrounding rigor, the part where an answer has to be reproducible and defensible. Frekil's design reads as an attempt to capture the first without giving up the second, keeping the model in the role of tireless drafter while the science stays inspectable.
The foundersFrom Stripe and Sony to epidemiology
Frekil was founded in 2025 by two IIT Bombay graduates. Nikhil Tiwari, the CEO, previously wrote software at Stripe, Amazon, and Marsh & McLennan. Shivesh Gupta, the CTO, worked on systems software at Sony in Japan. Neither came out of pharma, which is either the obvious risk or the point, depending on how you read it. Their background is in building infrastructure that has to be fast, correct, and load-bearing - the qualities a regulated evidence pipeline needs more than industry pedigree.
The company's operating principles carry the fingerprints of that engineering culture. Frekil lists five: customer success first, a bias for action (the founders promise support responses within 10 minutes, around the clock), scientific rigor with no methodological shortcuts, radical transparency in which every transformation and statistical choice is documented, and the data-privacy rule that started it all. For a two-person team selling into an industry that moves cautiously, promising a 10-minute human reply is less a feature than a strategy.
The businessSoftware instead of a study
Frekil sells software where the industry has historically bought projects. Instead of commissioning a study and waiting two or three quarters for a deliverable, a customer runs studies themselves on a platform that plugs into their existing data stack. That is the classic shape of a good software business - replace episodic, high-cost services with a tool people reach for repeatedly - applied to a corner of life sciences where the episodic, high-cost version has been the only option.
The company is early. It came through Y Combinator's X25 batch, working with partner Nicolas Dessaigne, and raised seed funding with participation from Y Combinator and 7BC Venture Capital. The team is tiny, the customer list is still forming, and the claims about speed will be tested by regulators and biostatisticians who are paid to be skeptical. But the thing Frekil has picked is real: an expensive, slow, mostly hidden process that sits between patient data and the decisions made from it.
There is a second-order effect worth naming. When a study takes a year and costs six figures, teams ration their curiosity. They ask only the questions they are fairly sure will pay off, and the softer questions - the hunches, the double-checks, the "what about this subgroup" - go unasked. Cheap, fast studies change that math. They make it affordable to be curious, which in a field built on catching signals early is not a small thing.
If the arrow holds - if months really do become minutes without giving up the rigor that makes evidence usable - the interesting part is not that a study gets cheaper. It is that a company might run ten studies where it used to run one, and ask questions about its drugs it previously could not afford to ask.