Tools for seeing inside neural networks.
JNDU Labs is an early-stage AI lab building interpretability tooling: software that helps researchers and engineers understand what their models have learned, and why they behave the way they do.
What we're building
Modern models work. It is much harder to say how.
Our first projects are small, focused tools for mechanistic interpretability, built in PyTorch and JAX and developed in the open where we can.
Feature discovery
Dictionary-learning pipelines that decompose a model's internal activations into features a person can read, name and search.
Circuit tracing
Tooling for following a single behaviour through a network, from input tokens to output, with interventions that test each step.
Shareable reports
Browser-based views that turn an analysis run into something a teammate can open, explore and question without rerunning it.
Company
Small, early, and in active development.
JNDU Labs was founded in 2026 by Jonathan Du. We are pre-launch: the work right now is experiments, prototypes and infrastructure, with a first public release planned for early 2027.
If you work on model internals, evaluation or safety and want to compare notes or try an early build, we would like to hear from you.
- Status
- Pre-launch, building the initial product
- Focus
- Interpretability tooling for neural networks
- Stack
- PyTorch, JAX, cloud GPU infrastructure
- Founded
- August 2026
- Location
- Mountain View, California
Contact
Interested in early access?
Research collaborations, early builds and general questions all go to the same inbox.
jonathan@jndu.me