Home›Companies›Perceptual Labs
Perceptual Labs

Building intelligence that can understand and rewrite itself

Active Founders
Kevin Miao
Kevin Miao
Founder/CEO
Co-founder at Perceptual Labs. Senior Research Engineer at Apple and Lecturer at UC Berkeley. Previously researcher at Berkeley AI Research (BAIR). MS and BA in Computer Science from UC Berkeley. Born and raised in the Netherlands; speak 8 languages!
Frankie Li
Frankie Li
Co-Founder
Co-Founder at Perpetual Labs. Previously helped train models and led product at Fundamental Research Labs (a16z-backed, MIT spinout), hire #3 where he scaled products to millions of users and revenue, including Shortcut (Excel copilot used by 1000+ companies). Harvard Grad.
Company Launches
Bryel - Building your in-house frontier AI lab
See original launch post

Hey YC, we’re on a mission to build a billion models, trained by you. We’re Kevin and Frankie, co-founders of Bryel.

TL;DR

We’re your cracked AI training team, on demand. We partner with your company, build your in-house model training capabilities, and leave behind Bryel Foundry, our software that makes model development and continual improvement 10x simpler. Think of it as your in-house AI lab. We’ve trained models for Apple, Framer, and F500s.

Ask: Considering training your AI model? Email frankie@bryel.ai

https://www.youtube.com/watch?v=LsVUbDIG2RY

Problem

When you train a model for everyone, you train it for no one. 

Frontier models are incredibly capable generalists, but they fall short on the specifics. So teams build harnesses to work around the edges and cover the gaps: more prompting, more harness infra, more retries, and stricter evals.

Eventually, quality plateaus. We’ve met many teams that are buried in harness work, layering on more and more infrastructure and patches to keep things working. That’s where model training is worth a serious look. With recent advancements in the open-source ecosystem and post-training methods, a few thousand good examples can produce a model that beats frontier APIs on your task, and it’s cheaper to run and deployable on-prem.

The challenge has always been that training is slow, expensive, experimental, and requires specialized knowledge. That’s where we help.

Solution

We forward deploy into your company, to help you set up your AI model training capabilities end-to-end, and hand off software to help you maintain and continually improve your AI models. Our product, the Bryel Foundry, has been built from everything we learned doing this by hand, many times over.

Models we commonly help train:

  • OCR/Classification Models that run 20x faster.
  • Search & Embedding models that cost 5x less. 
  • Long horizon agent models that match frontier performance at 5-8x lower cost

One more thing, we have a second ask. If you know companies that have a lot of proprietary data and no ML team (think insurers, logistics, healthcare, legal)? We'd love a 20-minute intro. We're not looking to sell, just to learn. Happy to return the favor.

uploaded image

Perceptual Labs
Founded:2026
Batch:Fall 2026
Team Size:2
Status:
Active
Location:San Francisco
Primary Partner:David Lieb