R·01
Interpretability at systems scale
Making the internals of large models observable in production, not just in notebooks — tracing behavior back to mechanism.
- Feature attribution
- Activation tooling
- Eval harnesses
an independent research lab
Webuildthetoolsintelligencewillneedtobeunderstood.
An independent lab working at the seam between machine-learning systems and the people who have to rely on them.
links
Mission
A lab, not a product company.
Northlight exists to do a narrow thing well: build the infrastructure that makes advanced machine-learning systems legible. Not another model, but the instruments around it — the ways we measure, interpret, and hold such systems accountable.
We work in small cycles, publish what we learn, and ship the tooling as we go. The bet is simple: the field is long on capability and short on understanding, and the gap is an engineering problem as much as a scientific one.
Research
Three threads, one aim: the instruments that make machine intelligence legible.
R·02
Agents that stay steerable
Control surfaces for long-horizon agents: how to keep intent, constraints, and human oversight intact as autonomy grows.
- Oversight loops
- Constraint solvers
- Trace analysis
R·03
Infrastructure for trust
The unglamorous plumbing — provenance, reproducibility, and measurement — that lets anyone verify a claim about a model.
- Provenance
- Reproducible evals
- Open tooling
Writing
People
Working on the same problems?
We collaborate with researchers, fund small tools, and hire rarely but deliberately. If the work above is yours too, get in touch.