pre-alpha · more coming soon
Plicara Labs
We are an applied research lab specialized in training and benchmarking AI models. We develop opinionated and thoughtfully constructed artifacts, and we like to share what we learn with the broader ML and AI community.
The mission
01. the mission
“Be a strong bedrock on which AI systems are thoughtfully evaluated and built.”
We are entering a phase in computer science and applied research where the most important quality an engineer or developer can have is great taste. Now that digital artifacts are easy to generate, the differentiator is knowing what matters, what doesn't, and how genuinely high quality work defines the things we build.
The point of this applied research lab is to be a good source for developing that taste: finding important patterns to work with, and building the models that underpin this ethos. We train models and provide benchmarks that are opinionated, carrying the design patterns we have found helpful and important in our own experience, and we like to explain the why of it all simply, so everyone can grow along with us.
Models
02. the range
Releases will be named after different paper plane designs. Nothing is released yet; more coming soon!
More coming soon. The range →
Tools
03. shipped
Each released artifact or tool gets its own repository,
Apache-2.0. regexbench and labloop are
published on PyPI; benchmark results are coming soon.
Evaluate generated regular expressions: semantic equivalence, correctness, and ReDoS safety.
Inspired by autoresearch, an agent-driven experiment loop: propose a change, run it time-boxed, keep it only if the metric improves.
Runs regexbench across models and publishes the
numbers: scores, methodology, and a re-run command.
Principles
04. how we work
Open weights
What we train, we release and explain. Weights you can download, run, and check. Black boxes are not trustworthy, nor in the spirit of this lab.
Reproducible research
Anything we publish, you can re-run and (we hope) understand.
Reliable systems
Boring infrastructure is a feature, and we take care to build systems that are robust and generally scalable.
Simple is not the enemy of powerful
We reach for the plainest thing that works, and our models, research, and benchmarks carry this ethos.
Contact
05. write us
Something here useful to you, or wrong? Open an issue, or write to us. We read everything.
info@plicara.ai