PerfectBit
ActiveTraining data for frontier AI labs
This company has previously operated under “Gikl, Inc”. A rename frequently marks a pivot in positioning or product — useful raw material for variant ideas.
About
We create a new kind of data for training AI models. Most LLMs are pre-trained on noisy web-scraped text, but they hallucinate and still fail on tasks that humans find trivial. PerfectBit creates high-quality training data that's correct by construction. We verify against physics simulators, scientific databases, formal proof systems. LLMs, robotics, AI for Science, and more.
Founders · 2
I worked as Director of Media Generation at Meta before 2026 for 11 years. I was managing the Media GenAI foundation model research and development, including efficient media generation, text to image generation (Emu), image editing, Movie gen, text to video, video editing and character consistent image and video generation. Previously, led efficient deep learning for computer vision teams supporting on-device models for AR/VR. I was Assistant Professor at Stanford University
Led teams in the Core Llama group at Meta Superintelligence Labs. Senior Staff Research Scientist across 9 years at Meta spanning LLM pre-training and post-training, inference optimization, full-duplex speech models, and computer vision vision models. Before tech: PhD in Physics, published in Proceedings of the National Academy of Sciences and Physical Review Letters, co-authored with Fields Medalist. Educated at Stanford, Rice, Columbia, post-doc at a National Lab.
Related startups
The API for real-world training data.
Frontier models for critical domains
