We build probabilistic AI/ML tools, programs, and teams for organizations working in humanity’s most important domains.
More than 80% of data science efforts are wasted. We believe this is caused by the uninterpretability and brittleness of traditional ML technologies, which forces inefficient data science practices. We take a different approach to ML using probabilistic methodology, which allows us to complete months of work — with plainly understandable, actionable results — in days.
We are terraformers
Our goal is to improve the conditions of your data science program, build a self-sustaining probabilistic ecosystem, and (mostly) move on.
Probabilistic AI/ML R&D
We have performed probabilistic programming R&D work for some of the world's biggest AI buyers working in the most high-risk domains including the US Department of Defense and multinational corporations.
Low risk through radical candor
R&D is risky — data science R&D doubly so. This is mainly because knowing the right question is hard, and modern ML tools make it as hard as possible to ask questions and to know which questions are answerable with your data. We use technology that allows us to determine quickly (usually in less than two weeks) whether your data can get you where you want to go, and if not, to determine how to address these knowledge gaps. You either learn from your data, or very quickly learn that your efforts are better spent elsewhere.
- Problem identification
- Onboard to your systems and data
- Determine whether what is feasible with the data.
- If data are sufficient, build out the project
- If data are insufficient, develop a data collection or active learning plan (optional)
Build value by building your people, not by bringing in consultants.
Building a probabilistic programming initiative is hard. We can help.
There is a reason probabilistic ML hasn't taken over: it's hard. It requires users to be skilled in Bayesian statistics, machine learning, and usually a domain. Hiring people with expertise in one of these areas is hard enough, but hiring people with them all is nearly impossible, especially if you do not already have people with the expertise to recruit and interview them.
We build probabilistic ML teams by improving the people you already have. We engage on a specific project and serve as PMs and leads for that project while training your personnel on the job. As the team needs to grow we can perform recruiting and interviewing services.
Why Probabilistic AI/ML?
We have chosen the Bayesian probabilistic AI/ML paradigm because it prioritizes understanding and uncertainty quantification, which is critical for the domains we work in. We have repeatedly observed first hand how traditional Machine Learning and Data Science methods produce models that cross validate well but fail in production for unknown reasons. In direct contrast to traditional ML, the Bayesian probabilistic paradigm provides the following advantages:
- Predictions come from distributions, which communicate variance and uncertainty
- Provides rigorous frameworks for choosing between models
- Does more with far less data
- Enables active learning, which helps determine which future data would help us best learn
Dedicated to Open Source and Rust
Since our whole shtick is building transparent and safe ML tools, it only makes sense to write performance-critical code in a memory-safe language like Rust, and to make the source available for scrutiny.
- Lace: Bayesian tabular data analysis for Rust and Python
- Changepoint: Bayesian change point analysis for Rust and Python
- rv: Probability distribution and statistical utilities for Rust
Get in touch
Redpoll is deliberate about the organizations with whom we choose to partner. If you’re interested in working with us, please fill out the brief form below and we’ll set up a time to connect!