Expand the hypothesis space of AI capability.
Frontier models keep gaining raw intelligence, but capability only matters when it can act: gather evidence, run analysis, follow a lead, and hold memory across engagements. We build agentic systems for domains where the work has to hold up — markets, elections, public trust, and science.
Today that means two instruments. Sentinel gives people and institutions a shared memory for information threats: detecting coordinated campaigns, attributing the actors behind them, and forecasting their next moves. Cato turns biomedical data analysis into a conversation: from dataset to defensible, cited, reviewer-ready results in minutes.
Agents need real environments.
Benchmarks are not the test. We learn what agentic systems can do by giving them consequential work — investigations, analyses, briefings — and studying where they hold and where they break.
Systems need memory.
One-off outputs evaporate. Whether tracking a threat actor or a research program, the instrument should compound: every engagement making the next one faster and better grounded.
Outputs should support decisions.
We build for briefings, alerts, figures, and findings that people act on. The point is not another surface to watch; it is to help someone decide.
The work is interdisciplinary.
Our products sit between research, engineering, intelligence analysis, statistics, medicine, policy, law, and institutional risk. We hire and build accordingly.
Researchers, engineers, and operators building shared instruments.
Our work spans data infrastructure, applied modeling, agent design, evaluation, and product systems — intelligence briefings on one side of the house, statistical analysis on the other. We care about systems that hold up under adversarial pressure and outputs that remain useful when decisions are urgent.
Research
We study how agentic systems reason, act, and fail — and how coordinated campaigns and complex datasets actually behave in the wild.
Engineering
We build the corpora, ingestion and execution systems, analyst tools, and evaluation harnesses that make reliable, recurring coverage possible.
Analysis
We turn evidence — actor behavior, source trails, statistical results — into briefings and findings that experts can verify and leaders can use.
Product
We translate the instruments into workflows people rely on: watchlists, alerts, recurring briefings, and conversational analysis.
How we work when the environment is uncertain.
Build the instrument.
The fields we work in need infrastructure, not one-off analysis. We build systems that accumulate memory and make the next investigation — or the next analysis — faster, clearer, and more grounded.
Stay close to evidence.
A claim is only useful when it is sourced, caveated, and falsifiable. Whether attributing an actor or reporting a statistic, we separate what is known, what is likely, and what remains uncertain.
Design for decisions.
The customer is usually under time pressure. Our work should clarify the situation and the options — the actor and the forecast, or the result and its caveats — without forcing another analyst workflow.
Handle sensitive work carefully.
We support people in active risk windows and work with data that demands care. Discretion, operational discipline, and careful communication are part of the product.
Prefer the simple thing that works.
We care about useful signal, not elaborate machinery for its own sake. If a direct briefing, figure, or alert solves the problem, that is the right output.
Work with us
Tell us what you are seeing and how we can help.
Get in touch