Why we are using plant breeding to test defense AI
Like the battlefield, nature is a complex and unpredictable system that crushes best-laid plans under constant and unprecedented change
Redpoll partners with Rutgers and DARPA to build AI that adapts to dynamic worlds
The OTACON project aims to deliver introspective AI
Announcing rv 0.8.0
A rust crate for building probabilistic programming tools
Make more informed decisions with fewer manual processes and less communication overhead. Based on human cognition research, Redpoll AI learns and stores its knowledge like people do, making it transparent and intuitive to both stakeholders and data scientists.
Learn the complex relationships within your unstructured data. Simulate or evaluate the likelihood of hypothetical events, predict new values given any number of conditions; quantify uncertainty -- all without needing to partition, transform, or fill in missing data.
No repeated cleaning, coding, or model validation. While the traditional approach to analytics is to model questions about data, Redpoll AI models the data itself, which streamlines the analytics cycle and eliminates the cost of asking new questions.
From exploration to production in minutes. End users and machines interact with the platform through the same API, which can be accessed through the web or dropped into your containerized environment.
Data QA & Sustainable Learning
Find bad data; don't fit to it
Most AI are designed to predict data, which causes them to fit to noise (overfitting). The Redpoll AI is designed to model data, so Redpoll AI can determine which data are odd, which allows you to identify data that do not make sense and understand why.
Add or modify data without retraining
Traditional AI must be completely retrained when new data come in or old data are modified; throwing away past learning and wasting all the time and money spent on that learning. Redpoll AI learns like a person, allowing it to train only on what is new or revised, saving time and money, and preserving learning.
Example: find and fix a data entry error
 rp = redpoll.Client("rp.mydomain.com")  rp.odd_data("hour of day", top=3) rank row hour of day ---- --- ----------- 0 453 239 1 24 24.0 2 501 4.0  rp.update_data( ... row=453, ... column="hour of day", ... value=23.9 ... )
A demo web client displaying a network of genomic similarity with respect to a disease phenotype. Your interface will differ to reflect your requirements.
G x E x Anything
Discover and characterize multi-locus, epistatic traits, and characterize the modifying effects of environment, management, and treatment. By mapping information flow through the data, the AI finds and characterizes complex structure in unstructured data.
From GWAS to Genome Editing
Predict or simulate phenotypes from genotypes, or simulate the genetics likely to produce desired phenotypes. All information the platform creates is reversible, which allows users to simulate, optimize, and compute probabilities bidirectionally.
Design experiments and data collection plans that are less costly and more effective. The AI can identify gaps in its knowledge and can request specific data to fill those gaps; and it can evaluate the information content of hypothetical data to help users avoid over-collecting, or collecting the wrong data.
AI / ML / Analytics Consulting
Use the AI technology we're developing to solve some of DARPA's hardest problems to solve yours
AI and analytics projects are a big risk. It's difficult to know what — if anything — your data can do for you. Projects begin then change shape as analysts run up against the limits of the data and machine learning technology. Often what you get is what can be done rather than what must be done. We are developing a new type of humanistic AI that solves the slowness of data science discovery and reacts to complex, dynamic data in complex, dynamic worlds. Know what is possible. Solve new problems, faster, safer, and more reliably.
What We Do
- Feasibility Studies
- Automated decision pipelines
- Human-in-the-loop systems
After a project has been identified, we conduct a feasibility study in which we determine whether your data will be effective at achieving your goals. If feasibility is confirmed, we plan, design, and build out the solution; otherwise we can develop a plan to improve the effectiveness of your data.
Get Results Months Sooner
Our in-house technology allows us to determine what is possible with your data and to discover transparent knowledge with unrivaled speed. By automating many of the processes that require data scientists months to complete, our tools allow us to complete in days tasks that would otherwise take months or years.
Safe, Reliable Models
We believe in safety from the bottom up. Our AI builds models that are robust to weirdness in the real world, and that can quantify uncertainty and ask for human intervention. And we develop our core AI in Rust: a safe, performant language that eliminates the cause of ~70% of security vulnerabilities.
We are engineers, scientists, psychologists, and AI researchers.
The AI industry is built around black box algorithms, like deep learning, which require incredible amounts of data, and which can fail in unexpected ways. These algorithms have found great success in domains for which data are plentiful and cheap, and for which the cost of failure is low -- domains like ad targeting, speech processing, and image classification. But in safety-intensive, scientific domains like biotech and public health, data are sparse and expensive, and the cost of failure is potentially lost lives. Using the same algorithms is unproductive, unsustainable, and potentially dangerous.
We believe the best way to solve these hard, impactful problems safely and sustainably is to start from the ground up. To develop new AI techniques based on human cognition. To build expressive, intuitive, and transparent AI that can be operated and understood by both data scientists and domain experts.
Baxter Eaves, PhD
Dr. Eaves has 10 years of experience building analytics tools and designing human cognitive systems that learn, teach, and selectively trust like people. He has applied these methods to problems in linguistics, image processing, biotech, and agriculture.
Patrick Shafto, PhD
Dr. Shafto is the Henry Rutgers Term Chair in Data Sciences at Rutgers University - Newark. He has led a number of projects for agencies like DARPA, DOD, and NSF; and his publications have appeared in top journals of machine and human learning.
Principal ML Engineer
Michael has 14 years of research and engineering experience. He has built production models for healthcare, agronomy, finance, and law; and has conducted research in the areas of high-energy physics, differential geometry, plasma physics, and high-performance computing.