Walk into an oncology practice, and before a single patient is seen, you’ll likely see somebody reading a chart. Maybe it’s a nurse pre-charting a complex case for a 9 am visit; she might also be pulling together imaging and genomic results and noting which regimens the patient stopped six months ago. Or perhaps you’d encounter a registrar working through a mandatory cancer-registry abstraction, or a research coordinator reading a chart to decide if a patient might match the inclusion criteria of an active trial.

A complex cancer patient’s record can run 300 pages across clinical notes, scanned documents, faxes, pathology reports, imaging results and genomic panels. The oncology coordinator’s job is to read all of it, hold a trial protocol with its inclusion and exclusion criteria in their head, and decide whether this patient might match for a trial—and then do it again for the next patient. Screening alone accounts for up to a quarter of coordinator time at an oncology site.

This is the main bottleneck in bringing life-saving cancer drugs to market. Eligible patients are sitting in waiting rooms, showing up to appointments and being documented across every EHR in the country. The problem is that someone needs to read the records.

The pressure on this bottleneck is accelerating. Oncology trials now represent 41% of all clinical trials globally, with trial starts up 58% over the past decade; over ten thousand oncology trials are actively recruiting right now. AI-assisted drug discovery is compressing the front end of the pipeline in ways the industry has not absorbed yet. All of this is happening as the WHO projects a 77% increase in new cancer cases by 2050.

Triomics* has built the oncology-specific AI infrastructure to solve this problem.

Why this has been unsolvable until now

People have tried to solve this before. The earlier attempts ran on keyword search and rule-based matching, and they did not work very well, because the job is not CTRL+F. When a protocol requires that a patient has “failed first-line chemotherapy,” you cannot keyword search for that. Generally you have to reconstruct a chronological treatment history from notes written by different physicians over years, link a specific regimen to a progression event documented months later in a radiology report, and make a clinical inference. The earlier tools could not do this with enough accuracy to earn the trust of physicians and research coordinators – and a tool that research coordinators do not trust is a tool that does not get used.

What has changed is the model layer: reasoning over long, messy, multimodal records. And this is what LLMs have gotten good at in the last two years.

Sarim and Raj

Model improvements unlocked the category, but getting it to work on an oncology chart well enough for a research coordinator at a premier institute to trust the output is a different task. This requires somebody who understands oncology workflows deeply enough to know what the right answer even looks like, and a person who can build production-grade AI on the kind of long, messy, multimodal data that an oncology record actually is.

Triomics has both founders: Sarim Khan was a published biotech researcher at MIT and worked closely with clinical research and care teams in the Boston ecosystem, while Raj Singh led generative AI research at Adobe with multiple publications spanning multimodal understanding and controllable text generation. One customer we spoke with called them a “deadly combination” for cancer.

What Triomics built

The end customer outcomes we’ve heard are only possible because of a foundational architectural decision. The team built OncoLLM, a family of eight oncology-specialized models ranging from 3B to 72B parameters, structured as an agentic reasoning system rather than a single model call. It runs on top of OncoIndexer, the company’s data pipeline with integrations into Epic, OncoEMR, and iKnowMed – the three systems that cover roughly 70-80% of the oncology providers. The output is a structured patient profile built from the full longitudinal record and evaluated in real time against trial eligibility criteria.

Accuracy on patient-to-trial matching has crossed 95%, which is the threshold cancer centers will actually trust. In an American Society of Clinical Oncology presentation, the Medical College of Wisconsin Cancer Center reported a greater-than-30% increase in trial enrollment across their active trial portfolio and a 67% reduction in screening time using Triomics. Memorial Sloan Kettering Cancer Center (MSK), MD Anderson, Yale Cancer Center and its partner Smilow Cancer Hospital, as well as Mount Sinai Tisch Cancer Center, are among the well-known oncology institutions working with the company.

Customers say the product saves the physician and coordinator time and drives revenue at the same time. ROI conversations at academic medical centers are not hard.

The future of Triomics

We are just as excited about what happens outside the academic medical center. About 85% of cancer patients are treated in community settings, and most community practices run few trials or none at all. The operational pain in community oncology is different – pre-charting complex patients before visits, curating real-world data for payer and pharma reporting, and completing the mandatory cancer registry abstractions every practice is required to submit. These are the workflows that are actually consuming staff hours in the community setting.

Because OncoLLM builds the structured patient profile once, every workflow – matching, pre-charting, registry, cohort visualization – pulls from the same profile without redundant integration or redundant compute. Triomics has landed in community practices on pre-charting and reporting pain, which has nothing to do with trials, and then turns that footprint into a trial screening network on behalf of academic sites and pharma sponsors who want eligible patients they cannot otherwise find.

The flywheel is intuitive: More sites generate more structured patient data. More structured data makes the network more valuable to sponsors, who today rely on CROs to manually survey sites and estimate patient pools, a process that takes months and is usually not precise. Sponsors can instead query a real-time, AI-structured feed of data across a national footprint. Site footprint compounds and pharma value compounds with it.

The oncology drug pipeline is producing more trials and therapies than the existing screening infrastructure can absorb. The accuracy required to put AI inside a clinical trial workflow only became possible recently. Sarim and Raj have built a harness that capitalizes on this advancement; they have the early customer set to prove that the world-class cancer centers will trust it.

What if we had real-time structured data on every cancer patient in the country? How many lives could be saved? How many therapies could come to market faster?

We are going to find out together. We’re excited to partner with the Triomics team and lead the company’s latest round of financing.