Research
Tensor Lab builds AI systems for people who know their field deeply.
You have spent years—often decades—learning how your domain actually works. You know when a diagnosis fits the patient in front of you. You know when a forecast ignores weather patterns that matter. You know when a schedule violates rules that are not written down in any manual. That knowledge is your moat, and it is the part AI cannot replace.
What AI can do now is the mechanical work: drafting, coding, scheduling, summarizing, predicting. It writes the first version. You decide whether that version is right.
This is how we think the future should work.
How We Build#
Workflows that wait for you#
We design systems where AI handles routine work and stops at the moments that need human judgment. A workflow might generate a draft report, but it will not send it until you have reviewed and approved it. An agent might triage a hundred cases overnight, but it will surface the five it is uncertain about for your morning review. The system works continuously; the decisions stay with you.
Agents you can correct#
Our agents do not just produce output and move on. They expose their reasoning so you can see where they went wrong. When you correct them—“this patient needs a different protocol because of their history” or “this forecast is off because it missed the upstream dam release”—the system learns. Your corrections become part of how the agent works next time, not lost feedback sent to a black box.
Interfaces that put you in control#
We build interfaces where your “no” matters as much as the machine’s “yes.” You can override any recommendation. You can flag any output as wrong. You can define the rules the agent should follow. The interface is designed around the assumption that you know more than the system about the specific case in front of you, because you usually do.
What We Work On#
Tab Recorder#
Tab Recorder captures and transcribes research conversations, interviews, and field notes, then turns them into structured summaries. It is built for people whose work happens in conversation—clinicians, researchers, investigators. You talk. It types. You review the summary, fix what it got wrong, and move on. The knowledge stays accurate because you checked it.
Machine Learning Foundations#
We work on the underlying technology that makes AI more capable and reliable. But capability is not the only goal. We ask: can a person with ten years in a field look at what this model produced and recognize whether it is correct? A model that looks good on a test but falls apart under expert review is not useful to you. We aim for systems that earn your trust because they get the details right that only an expert would notice.
AI Safety and Alignment#
AI safety, in practice, means making sure the system does not quietly go wrong in ways you would only catch too late. We build systems that explain their recommendations in terms you can evaluate against your own knowledge. When an AI suggests a treatment plan, you should be able to see the chain of reasoning and say, “That step would not work for this patient.” Safety is legibility. Safety is making sure the person with ground truth can still spot the error.
Applied AI for Public Good#
In healthcare, we build tools where the clinician controls whether a recommendation is used. In education, we build systems where the teacher decides whether an adaptive lesson sequence matches how their students actually learn. In environmental work, we build models where the field technician can override a prediction based on local conditions the data never captured. The AI generates possibilities. You choose which ones are real.
Open Research Infrastructure#
We build and share free tools that let domain experts use AI without hiring a software team. Our goal is to lower the barrier so that the people with the knowledge—not just the people with the engineering budget—can build, adapt, and control the systems that affect their work.
Working With Us#
We collaborate with universities, nonprofits, and research groups, and we are especially interested in partners who hold deep knowledge of the fields where AI is being applied. If you know your domain well and want tools that serve your judgment instead of replacing it, we want to hear from you.