Frekil
Frekil builds AI agents that run observational real world evidence studies for pharmaceutical, biotech and contract research teams on data they already hold or license, from protocol through final report. A researcher states a question in plain language, and an eight stage pipeline moves it through literature review, question clarification, cohort building, a causal diagram, a statistical analysis plan, data extraction, execution of that plan and report writing. The protocol and analysis plan are locked before any data is touched, researchers can review or change each step, and the AI writes R, Python and SQL code rather than results. The work is sold as Frekil Protocol for study design, Frekil Insights for feasibility counts and cohort checks, and Frekil Analytics for tables, figures and reports.
Use cases include trial feasibility, external comparators for single arm studies, post authorization safety studies, health economics and EU joint clinical assessment evidence, label expansion and comparative effectiveness. Protocols are checked against ICH M14, ENCePP and FDA real world evidence guidance, and reports against STROBE, RECORD, TARGET and the EMA study report template.
Frekil deploys in the customer's own cloud, on Databricks, Snowflake, AWS, Google Cloud or Azure, with EHR, claims, registry and licensed datasets such as MarketScan and Optum. Its models never see patient level data, and approved code runs in a separate sandbox with every change and sign off logged. Nikhil Tiwari and Shivesh Gupta founded it in San Francisco in 2025, and it went through Y Combinator's Spring 2025 batch. Pricing is per study or by platform, on quote.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
Language model agents do the study work. They clarify the research question, review literature, define the cohort, build a causal diagram, draft the statistical analysis plan, write extraction and analysis code and assemble the report, which is the labor a real world evidence team would otherwise put together by hand.
The statistics themselves run deterministically: the agents' code executes in a separate environment, and a biostatistician can edit it before it runs. A team could write its own scripts, but what it buys from Frekil is the drafting and coordination across all eight stages, which the company sets against the 6 to 18 months it says a conventional study takes.
Review is built into every stage. Researchers responsible for the protocol, cohort, plan and results each sign off, biostatisticians can revise any proposed script, and the protocol and statistical analysis plan are locked before execution. Changes and sign offs are recorded with identity, time and reason.
The model drafts the analysis logic while separate computation produces the numbers, so a team can read the code and compare the output with the report before approving it. Sign off covers both the setup that guides the agents and the finished study, and the company frames the product as support for scientists rather than a replacement.
The pipeline's eight agents are named by function, from literature review and question clarification to a causal diagram builder, analysis plan generator, data extractor, plan executor and report writer. Their artifacts are reviewable Python, R and SQL, data snapshots and software packages are versioned for reruns, and the report traces each claim to the code, data analysis or literature behind it.
The architecture and its outputs are open to inspection. The models behind the agents are not identified, and no release notes, evaluation sets or performance thresholds are given for protocol drafting and code generation.
The study workflow is Frekil's own, with language models writing code and a separate environment running it. Databricks, Snowflake, AWS, Google Cloud and Azure are named as places to connect data or deploy the software, and the statistical libraries used in an analysis are visible in each study's artifacts.
The upstream provider of the language models that draft protocols and code is not named, and neither are base models, versions, training provenance or where each model is hosted. A customer can see every package that touched its data but not which outside model wrote the code.
Worked examples are the main evidence. One sample study of 1,856 ICU stays reports a weighted hazard ratio of 0.59 with an E value of 2.80, and a sample design passes 26 of 26 checks with a causal graph that flags two unmeasured factors. Three testimonials, from a senior biostatistician, an epidemiology professor and a clinical researcher, are anonymous.
The claim of studies in minutes to hours rather than months is tied to no named sponsor, observed study or controlled comparison. No evaluation shows how often generated protocols or code pass clinical and statistical review, and no named deployment reports a measured result.
The design keeps patient data away from the models. The language model that writes code never sees, processes or trains on patient level records, the customer approves code before it runs, and execution happens in a separate sandbox inside the customer's own cloud. Transformations, data versions and code versions are logged for review, and the privacy policy covers personal information in the wider service.
Retention, deletion and training use are described less fully for the other content the models do see, such as research questions, internal protocols, literature and feedback on generated outputs.
Frekil says its generative models work without viewing patient level data and that study computation runs inside the customer's virtual private cloud, an architecture built to limit what leaves the customer's environment. The site states HIPAA and GDPR compliance, and the privacy policy covers general personal data.
Whether the company acts as a business associate when its software processes identifiable records in a customer's environment is not said. No business associate agreement, eligible service tier or obligation on downstream providers appears in the public terms, so the separation of model and patient data is the stated protection.
The trust page describes isolated research execution, access boundaries, a code and data audit trail with who, when and why for every change, and version records for reproducing study output, along with electronic record controls aligned with Part 11. Each customer keeps its own cloud and data access controls.
Those are product features rather than an independent attestation. No SOC 2 report, accredited ISO certificate or outside security assessment is named, and no audit scope is given for the hosted parts of the code generation service, which is where Frekil itself holds customer research content.
Frekil is research software for observational studies on existing data and does not run interventional trials, recruit patients or operate sites. Protocols are checked against FDA real world evidence guidance, ICH M14 and ENCePP, reports follow the HARPER protocol template and the EMA final study report template, and electronic record controls are aligned with 21 CFR Part 11.
Those checks concern study documentation and methods. They do not mean a regulator has accepted a study design, cleared the software or approved a submission built with it, and the product has no diagnostic or treatment function for individual patients.
The workflow guards against bias in the evidence. It records approvals, locks the analysis plan before data access, flags unmeasured factors in the causal diagram and lets epidemiologists test confounding with positive and negative control outcomes, sensitivity analyses and E values. Reports map their documentation to named reporting standards such as STROBE, RECORD and TARGET.
Those controls help researchers judge bias in their own cohort. Frekil gives no subgroup performance results for its language models and no published criteria for validating agent updates, so how reliably the agents draft correct designs across study types is a separate question the audit trail does not answer.
Errors can be traced and rerun. Analysts inspect the generated code, rerun the analysis and follow any result back to the data snapshot, code and package versions that produced it, and every decision carries who, when and why. Product questions go to the founders' contact address.
The site terms carry general disclaimers and liability limits, and enterprise purchasing is negotiated separately. No accuracy warranty covers generated protocols or code, no service commitment covers correcting an analysis, and no indemnity applies if a regulator rejects a study built with the platform.
Frekil reads EHR, claims, registry and pharmacy data the customer already has, including licensed sets such as MarketScan and Optum, in native schemas with no common data model required. It works where records already sit, in Databricks, Snowflake or a customer warehouse, handles ICD-10, SNOMED, RxNorm and ATC coding, and can reuse a customer's own code and analysis libraries.
Outputs are study artifacts, not chart updates, so there is no write path into a clinical record. No named customer is shown moving data through the product from a specific EHR or warehouse.
Approved code runs in a sandbox inside the customer's own virtual private cloud, beside the data the research organization already holds. Databricks, Snowflake, AWS, Google Cloud, Azure and other data warehouses are supported, patient level data stays in the customer's environment, and the trust page lists region specific data residency controls, which matters for studies across countries with local data rules.
The model that drafts study logic needs no access to clinical records. Where the model endpoint runs, and where investigator prompts and service metadata are stored, are not specified by region.
Two ways to buy are offered: per study, which gives a clear unit of purchase, or a platform arrangement for teams planning several studies. The work can also be bought in parts, as Frekil Protocol, Insights or Analytics, and the company meets prospects at conferences such as ISPOR Europe 2026.
No per study price, annual fee, minimum purchase or charge for extra reviewers is published. The customer supplies data it owns or licenses, and integration, a different cloud setup or dedicated biostatistical support can change what a quote includes.
Pharmaceutical companies, biotech teams, contract research organizations and academic investigators with access to patient level data are the buyers. Frekil names oncology, cardiology, immunology and rare disease, and studies across countries that need local data handling. Solution pages cover post authorization safety and signal follow up, health economics and EU joint clinical assessment, label expansion, competitive evidence and trial feasibility.
The same workflow runs on many types of claims and record data, depending on what the customer holds. No deployments are named in any therapeutic area, and performance has not been shown for a particular specialty, study design or population.
Alternatives to Frekil
Other products that do the same kind of work as Frekil, and what each one does.
Aetion makes the Aetion Evidence Platform, software for designing and running real-world evidence studies on health data.
PhaseV provides clinical development software for pharmaceutical sponsors, biotechnology companies and contract research organizations, from offices in the Boston area and Tel Aviv, and says more than 40 biopharma companies and CROs use it.
Komodo Health sells real world data and AI analytics to drug makers, payers, providers and the consultants and investors who serve them.
Peer AI provides regulatory document authoring software for pharmaceutical companies, biotechnology developers, and contract research organizations.
OM1 is a Boston real world data and AI company that Richard Gliklich, a physician, founded in 2015. It sells to drug and diagnostics makers, health systems and payers, and the companies it names include Amgen, AbbVie, GSK, Johnson & Johnson, Hologic and Exact Sciences.
Mendel, based in San Jose and backed by Oak HC/FT and DCM, sells Hypercube, a platform that reads unstructured patient records for chart review, cohort building and clinical trial prescreening.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Per study or platform quote
|
The company offers a per study purchase or platform access, and the work can be bought as Frekil Protocol, Insights or Analytics. Rates, minimums and platform billing periods are not published. | Patient level data is processed within the customer's environment. No business associate agreement or tier is stated. | The software is deployed in the customer's virtual private cloud and connects to the customer's data. No setup or connection charge is listed. | Vendor Published |
Customers supply data they own or license. The published offer does not give separate rates for study volume, collaboration or cloud support.