RESEARCH

RA1 Labs pursues two tightly-coupled research directions, both resting on one bet: systems-level understanding of neural computation unlocks advances that surface-level approaches cannot reach. We build from first principles and verify before we claim.

Inference Architecture

Hardware-level optimization of neural inference. We design custom CUDA kernels, dispatch systems, and execution primitives from scratch, targeting real deployment constraints rather than theoretical ones — inference latency, memory hierarchy, numerical stability, and behavioral fidelity across architectures.

This direction is embodied in Raven, our GPU inference engine. Raven: verify-all parity at 0.00% divergence across 8 models.

Mechanistic Interpretability

Behavioral analysis of neural systems at the mechanistic level — understanding what models actually compute, not what their outputs suggest. Our focus: behavioral attribution through hardware traces, circuit-level analysis of model decisions, and verification infrastructure that answers a single question — does this model do what we think it does? More on our interpretability work.

Method

• First principles — build the solution from the ground up when existing tools don't fit.
• Verification-first — prove behavior before publishing.
• Compression as proof — what compresses well is what matters.

RA1 Labs operates as a US-registered LLC, founded 2026 by Kanishk Joshi. Independent, self-directed, no external funding.