Octozi

Models drive the core work, on a data workspace that carries some value without them. The company sells AI agents across five modules (data management, clinical science, safety, medical monitoring and clinical operations) that detect discrepancies, draft queries and explain flagged items, using fine-tuned language models with heuristic algorithms.

Underneath is a unified data model that harmonizes EDC, safety, laboratory, CTMS, eCOA and IRT data into one review workspace with source traceability, which a study team could use without the agents. The discrepancy detection and drafted queries do not exist without the models.

Reviewers decide every query; what the agents may do on their own is not set out. The company describes a human-in-the-loop design in which the platform flags discrepancies with a narrative rationale and drafts queries, and reviewers can modify or reject each suggestion before anything reaches a site; deterministic algorithms handle the parts that must be reproducible.

The five modules extend agents into safety review, medical monitoring and clinical operations, where the limits of what an agent does unattended, the confidence at which a flag goes to a person, and how an agent's action is marked in the audit trail are not published.

The base model and the approach are named; versions and update practice are not. The company's preprint says its discrepancy detection uses proprietary language models derived from Llama 4 and fine-tuned for clinical trial data cleaning, combined with heuristic algorithms built with expert medical reviewers, over a unified data model with source traceability and a narrative view explaining each flag. It reports recall and precision for the detector. What data fine-tuned the models, how versions are managed, and how a sponsor learns of an update inside a validated study are not published.

The model layer is the company's own, fine-tuned from a named open-weight base, with a training exclusion. The preprint names Llama 4 as the base of its proprietary fine-tuned models, and the home page says customer data is never used to train or update them and is stored in per-customer siloed environments. That makes the model chain short. The cloud host, any other sub-processors, and retention terms are not published.

Published company research with its method, short of an independent or live-study evaluation. The preprint (arXiv 2508.05519, August 2025, Octozi authors) ran 10 experienced medical reviewers on synthetic patients built from a phase 3 oncology database (51 case report form datasets, more than 150 patients) with seeded discrepancies: assisted review was 6.03 times faster, errors fell from 54.7 to 8.5 percent and false positive queries fell 15.5-fold, with an economic model of a representative phase 3 trial estimating $5.1 million saved, mostly from earlier database lock. The home page cites these as an 85 percent reduction in data errors and six-fold median throughput. No named customer deployment with figures is published; Debiopharm's venture fund is an investor.

Categorical commitments are published; retention is not. The company states that it never uses customers' private data to train or update its models, stores each customer's data in siloed environments isolated from others, encrypts data in storage and transit, gives customers control and visibility over access and usage, and uses zero-trust, least-privilege access with strong authentication.

The training exclusion matters for sponsors who will not expose unblinded trial data to a vendor that might learn from it. No retention period, end-of-study handling or privacy policy is published, and nothing is published on safety engineering for the language models beyond the heuristic checks paired with them.

Service commitments are published; no status and no privacy policy. The home page commits that customer data is never used to train models, is isolated per customer, encrypted and access-controlled, and lists GDPR compliance, but no privacy policy, data processing agreement or statement on HIPAA or business associate status is published.

Most trial data reaches the company from sponsors under subject identifiers rather than from provider records, where data protection law and sponsor agreements usually govern; the company does not say so itself.

A recognized attestation is named without the report. The home page's security and compliance section lists SOC 2 Type II and says the company undergoes regular audits and third-party vulnerability testing, with zero-trust, least-privilege access, strong authentication and encryption in storage and transit. No report, period, scope or trust portal is published. The company handles pre-submission trial data for sponsors.

No device claim is made, and the platform is scoped to trial data operations. Its output feeds datasets behind regulatory submissions, so electronic records and data integrity rules apply. The home page now lists 21 CFR Part 11 and "GxP Validated" among its compliance marks, and the preprint says the platform preserves source traceability; no validation summary, audit trail description or statement of how a model-drafted query is attributed is published.

Evaluation is published on a synthetic dataset, with nothing on where the detector flags. The company's preprint (Purri, Patel and Deurell, arXiv, August 2025; all Octozi authors) reports its discrepancy detector at 97.5 percent recall and 77.2 percent precision on an annotated synthetic dataset derived from a phase 3 oncology trial, and a reviewer study on the same kind of data.

Testing against discrepancies its authors seeded measures the detector against the company's own model of what goes wrong, not against the errors nobody anticipated. The risk that matters for automated data review is where the flags fall: a detector can be accurate overall while over-querying some sites, forms, source systems or therapeutic areas, and that pattern shapes the dataset a regulator sees. No flag rates by site or source and no governance framework are published.

The company publishes its detector's error rates with method, and the human baseline. Its preprint (arXiv 2508.05519, August 2025, Octozi authors) reports recall of 97.5 percent and precision of 77.2 percent (83.6 percent accuracy) for discrepancy detection on an annotated synthetic dataset built from a phase 3 oncology trial, with 10 percent of data points seeded with discrepancies.

In a study of 10 medical reviewers on that data, assisted reviewers' errors fell from 54.7 to 8.5 percent, precision rose from 44.6 to 93.2 percent and recall held (85.2 against 88.9 percent). These are company-run results on synthetic data, not performance on live trials, and the preprint is not peer reviewed. Reviewers can modify or reject every suggested query. No warranty, indemnity or remediation commitment is published.

No clinical record integration, and the trial systems it connects to are described only by type. For trial software, those are the systems that count. The platform harmonizes EDC, safety, laboratory, CTMS, eCOA and IRT data into one model with source traceability, and the company's positioning is to work with the systems study teams already use. No data capture or safety product is named, and no connector list or interface documentation is published.

Tenant isolation is stated; hosting and residency are not. The company says each customer's data sits in a siloed environment isolated from other customers' data, which matters when competing sponsors use the same vendor. Where the platform is hosted, in which regions, and whether a sponsor can keep data in a jurisdiction for a multinational trial are not published, nor is where the fine-tuned models run.

No price is published, and the value case is framed against a cost sponsors know. The company's economic model of a representative phase 3 oncology trial estimates $5.1 million saved per study, mostly from a five-day-earlier database lock, which a sponsor can compare with its own delay costs; it models a representative trial rather than measuring one. Sales are to sponsors and contract research organizations by demo request.

Coverage is stated clearly, with nothing validating it on live studies yet. The platform covers the clinical trial data layer (data management, clinical science review, safety review and reconciliation, medical monitoring and clinical operations) for pharmaceutical, biotech and medtech sponsors and contract research organizations. The only published evaluation used synthetic data built from one phase 3 oncology trial.