MedNova

Getting a drug to trial means reconciling thousands of documents against a wall of regulation, and in pharma a single fabricated line can set the timeline back months.

AmiaAI runs the regulatory submission on fully private infrastructure, turning weeks of manual gap analysis into a grounded, citation-backed draft across every eCTD module, with zero tolerance for hallucination built in.

The situation

Before a new drug can be tested in people, its maker has to convince the FDA through an Investigational New Drug application, or IND, the filing that must clear before human trials can begin. That submission pulls together thousands of documents, many hundreds of pages long, each of which has to line up with a dense body of FDA and international (ICH) guidelines. Teams do this by hand, reading the corpus and checking it against the requirements to find what is missing, work that takes weeks. Any error or omission can delay approval, and delay is enormously expensive. Because this is pharma, there is also no room for a system that invents a fact or a citation. The stakes make the slow, manual way the only trusted way.

Why it was hard

The obvious idea is to feed the documents to a large language model and ask. Two things break that immediately. First, scale. The corpus for a single submission far exceeds what a standard model can hold in context at once, so the reasoning has to work across far more text than the model can see in one pass. Second, trust. A regulatory submission has zero tolerance for hallucination, and every claim must be traceable to its source for audit, so a model that produces fluent but unsupported text is not merely unhelpful, it is dangerous. On top of both, the data cannot leave the building. Clinical and proprietary data rules out sending anything to a public cloud model, the very tool that would otherwise have the horsepower. Solving scale and trust at once, on private infrastructure, is the real problem.

The approach

AmiaAI runs entirely on fully private infrastructure: private models, private data, private deployment, so no clinical or proprietary information ever leaves the sponsor's control. To handle the scale, it uses a long-document reasoning approach based on TheAgentic Reasoning Workbench, that works across a document corpus far larger than a standard model's context window, so a full submission's worth of material can be reasoned over rather than truncated. The system runs the gap analysis, checking the corpus against Pre-IND templates and the eCTD structure (the standardized modules every submission must follow) to surface exactly what is missing, then generates IND content across the modules. Reliable retrieval across that corpus is also handled by the reasoning workbench engine, so every generated line is grounded in and cited back to its source document, which is what makes the output auditable and keeps hallucination out. Compliance is not a feature bolted on. It is the constraint the whole system is built around.

What happened

The change shows up in both speed and trust. Gap analysis that used to take teams weeks now produces a grounded first pass in a fraction of the time, with each finding and each generated section tied back to the source it came from, so a reviewer can verify rather than take it on faith. Because it all runs on private infrastructure, sponsors get that speed without handing sensitive data to an outside model. AmiaAI (formerly MedNova) is deployed with design partners across biotech, pharma, and regulatory consultancies, in environments where compliance is not optional and a wrong or unsupported claim is unacceptable.

What this means for you

If your teams prepare regulatory submissions, you already know the bottleneck is not expertise. It is the weeks of manual synthesis and gap analysis standing between your data and a submission-ready dossier, plus the hard rule that none of it can be fabricated or leave your control. That is what AmiaAI compresses, safely and on your own infrastructure, and it is the shape of what becomes possible when deep regulatory expertise and agent infrastructure are built together.