Products

Extract each claim and each step.
Run deterministic checks: dimensions, symbolic equivalence, limits and bounds.
Test with numerical experiments.
Have independent models and agents try to disprove it, including a cross-family judge that reviews methods as well as execution.

Every claim will be categorized as verified, partial, unsupported or refuted, with the evidence behind it. You can plug in your own checks too.

Check your AI's work

Today: math and citation checks available in Theo, and a cross-family judge can run in your workflow

With your team: checks built for your methods, run in your workflow. Spend less time reviewing.

Next: continuous review of every key result, and machine-checked proofs.

what we catch

Math and physics errors

A sign error or a units mismatch that changes the final answer.

Missing evidence

Claims with no supporting source, or a key source left out.

Citation errors

Made-up references, or papers cited for things they don’t say.

Unsupported claims

A statement presented as fact with no source or calculation behind it.

Tests that can’t fail

A check that passes no matter what the code does.

Reporting mistakes

A summary that says a check passed when it actually failed.

These are real errors we found, in Theo’s citation checks and in test runs where a model from another company reviewed research written by Claude.

Proof points

EMNLP 2026

CiteCheck, our citation verifier, accepted at the GroundLM workshop.

19 blocking errors

That would have invalidated later work, among 214 findings in 10 recent runs.

Early access

TheoOntology

Everything your lab has done, searchable by words or equations, for you and your agents.

We built our own math embeddings, and they have the highest published precision on ARQMath-3, the standard benchmark for equation search. This extends to figures, plots and more.

For your research, we connect your papers, reports, notebooks and data into one map of everything you’ve done. You and your agents can search all of it.

This mapping is also what makes the AI models your team uses better at your work. Before a model answers, it gets the right pieces of your lab’s history, found through math and natural language. In our own research, an open model solved 25% more problems when our search gave it a relevant paper to read first. We call this scientific RAG.

Try it on your lab’s work

Today: math-aware search across open scientific literature.

With your team: search and source-linked answers over your papers, reports, notebooks and raw data.

Scientific search for your lab

Proof points

0.74

Precision at 10 on ARQMath-3 equation search across 9 million formulas, the highest published as of October 2026.

3M + 125M

Abstracts and equations indexed for search across arXiv.

Early access

TheoConductor

Get the best from every model, on every task, for less.

Conductor breaks a task into steps and sends each one to the model that gives the best value for that step, whether that’s a frontier model, one of ours or a smaller open model. It works for quick questions but shines on long research runs.

Compare your model costs

Today: multi-model runs inside Theo.

With your team: we tune Conductor on your own tasks, then show you what it costs and how its answers compare with the best single model you use today.

Proof points

1 API

Works with OpenAI and OpenRouter, so your existing agents and scripts don’t need changes.

Major labs

Use models from Anthropic, OpenAI and Google alongside open models and our own, all in the same task.

Keeps running

If one provider goes down or hits a usage limit, Conductor moves the work to another.

Early access

TheoDiscovery

Find the next question worth asking.

TheoDiscovery runs on two engines. Serendipity reads the literature for methods from other fields that could move your problem forward. TheoX, a mathematical search engine that runs without an LLM, proposes novel conjectures from your data and tests them against counterexamples.

TheoDiscovery was used in a recent quantum physics paper with Gerard Milburn.

Try Discovery on your work

Today: Serendipity searches all of arXiv, and TheoX looks for new conjectures in your data.

With your team: We set up a custom implementation of Discovery on your data and problems, and you decide which ideas are truly novel and worth pursuing.

Serendipity: research directions from any paper

Give it a paper, and Serendipity returns ranked research directions grounded in a real paper from a field you may never have read by matching the structure of your equations, text and figures, then filters for ideas that hold up and are new. It comes up with cross-domain ideas more consistently than frontier models, whose suggestions tend to cluster around the same few.

87, 000

papers retrieved

12, 146

connections

2, 283

ideas that pass a novelty check

480

ranked research proposals

Connect people too

In development

In a large organization, not everyone can know everyone. Serendipity can match researchers whose work shares the same math, across teams and campuses, and suggest a problem they could take on together.

Saloons were where people met and traded ideas informally. When Prohibition closed them, US counties that lost them filed 8 to 18% fewer patents a year than counties that were already dry.

Who talks to whom shouldn’t depend on who sits near whom. Serendipity can match the researchers by the math they work on.

Open problems become benchmarks

In development

When someone solves one of the generated conjectures, we check the answer. Proofs go through Lean 4, and a program confirms counterexamples. An open problem can become a test question with a verified answer that you can run your models against. We keep these private so they stay out of training data.

who its for

Research teams

A discovery environment around your own data, definitions and examples. Each project can grow into reusable and personalized research software.

Courses

Every student creates a research program of their own and finishes the semester with software that captures all of their conjectures and solutions.

Model builders

Fresh open problems your models have never seen, with verified answers.

Proof points

187 vs 47

Different research ideas in 200 Serendipity results. Asking a frontier model 200 times gave 47.

Lean 4

Every mathematical proof is checked.

Hundreds

Open problems in our collection that no frontier model has solved yet.

Design partners

TheoIntelligence

Specialist models, trained and evaluated on your task.

We take an open model and train it on examples of your work, using the science training data we’ve built up. The model is yours, and it can run on our servers or yours.

See if a custom model would help

Today: our specialist models in Theo, and a 9B model with gains on GPQA, HLE and AIME.

With your team: You pick the tasks, and we show how your model does against a general-purpose model given the same information.

Proof points

8B to 1T

Parameters we post-train, across the Qwen, Nemotron, DeepSeek, Kimi and GLM families.

SFT + RL

Fine-tuning and reinforcement learning, with our own verification

Yours

You own the model and the data behind it.

How we work with your team

Most teams we join already have their own AI tools, consultants or model contracts. We work alongside them.

Scientists do the science

Physicists and mathematicians work the problem with your team, alongside our engineers.

Every result comes checked

Each step carries its sources and checks, so your experts can review the evidence quickly.

Measured against what you have

We agree on the tasks, baseline and success criteria with you, starting from your current tools and team.

Forward deployed

How we do it: a bootcamp on your problem

Bring one problem and a sample of your data. In two to three days our scientists and engineers build on it with your team, using whichever products fit. You leave with a focused prototype and a scoped plan for the pilot.

Book a bootcamp

01

1 hour

Scoping call

We hear what you’ve tried and what’s still hard, then pick a workflow for next steps.

02

2 to 3 days

Bootcamp

Our team works hands on with yours, on your data, with the products that fit the problem.

03

4 to 12 weeks

Pilot

We agree on the tasks, baseline and success criteria together. A forward deployed team then builds it into your setup, with weekly demos and a note on what we can reuse in your next project.

04

Ongoing

Deploy and grow

We run it in your environment, license what works and add the next product when you need it.

Who you work with

A solutions lead

Owns the scope, the timeline and the line to your team.

Domain scientists

Physicists and mathematicians matched to your problem.

ML engineers

Post-training, evals and the model side of the work.

Research partners

Your day-to-day contact for questions and feedback.

Deployed where you work

All of our products run on Theo, our research environment for physics and math. You can use them on our servers, in your own cloud account or on your own hardware.We don’t train shared models on your data. Anything built from it, like a custom model, stays with you. What we keep are the general methods and tools we bring to every project.

See how Theo works

Managed by us

The fastest start: Theo and our API, run by our team.

In your cloud

Your own cloud account, your keys and your choice of models.

On your hardware

On-premises and air-gapped deployments for enterprise engagements.

Ways to work together

1

Use Theo

Researchers start free with 1,500 credits.

2

Plug in

Connect your models, codes and data sources to Theo.

3

License

Run a custom version of Theo on your own stack.

4

Custom build

Our forward deployed team builds it into your workflow.

5

Discovery Partnership

You fund a research problem and our scientists work on it with yours.

Consultancies and integrators: bring our products into your client engagements.

Research partnerships

Paid research pilots

Pick one research question and we’ll work on it with you.

Guided by domain-specialized AI

Theo's models are specifically trained to excel at scientific research.

Structure before you execute

Build clear, complete research plans with guided validation (including questions, assumptions, methods, evaluation).

Steer research as it unfolds

Intervene at any step - adjust variables, assumptions, or methods throughout your research.

Re-run only what matters

Modify part of your research workflow without restarting everything.

See how results were produced

Inspect each stage of reasoning and execution, not just the final output.

Catch gaps early

Surface missing assumptions, unclear metrics, or weak methodology before spending any time or compute resources.

Tell us what you are working on

We start by learning your problem, then suggest the product, bootcamp or scope that fits it.

Thank you! Our team has received your request and will reach out within 24 hours.

Oops! Something went wrong while submitting the form.

Guided by domain-specialized AI

Structure before you execute

Steer research as it unfolds

Re-run only what matters

See how results were produced

Catch gaps early

Backed by domain-specialized AI for science

Observe & Research

Deep Literature Search

Question Formulator

Evaluator

Hypothesize & Solve

Theo Specialized AI Models

Modular Tools

Test & Analyze

Output & Feedback

DRO Output System

Feedback from Theo Collaborators Program

Feedback from RLHF