RA1 Labs is an independent AI research organization founded in 2026, working at the intersection of mechanistic interpretability and hardware-integrated inference. We operate as a US-registered LLC, building from first principles — no institutional backing, no external funding dependencies.
Core thesis: understanding how neural systems actually work — at the hardware level, through computation, at the model layer — is prerequisite to building trustworthy AI. We verify findings before publishing. We treat compression as proof.
Two core directions. Both rest on the same foundational bet: systems-level understanding of neural computation enables advances surface-level approaches cannot reach.
Hardware-level optimization of neural inference systems. We design custom kernels, dispatch systems, and execution primitives from scratch to handle real deployment constraints — not theoretical ones.
Current focus: kernel design targeting inference latency and memory efficiency in resource-constrained environments, with emphasis on numerical stability and behavioral fidelity across architectures.
Behavioral analysis of neural systems at the mechanistic level. Understanding what models actually compute — not what they claim to compute, and not what their outputs suggest they are doing.
Current focus: behavioral attribution through hardware traces, circuit-level analysis of model decisions, and verification infrastructure that answers — does this model do what we think it does?
AI is deployed at scale without real understanding of what these systems compute. You cannot trust what you do not understand. You cannot responsibly deploy what you cannot verify.
RA1 Labs exists to close this gap — to build the interpretability and verification infrastructure that makes trustworthy AI deployment possible. This requires work at multiple levels simultaneously: the hardware level, the architectural level, and the verification level. Most organizations choose one. We work on all three.
Active development across interpretability infrastructure and hardware-integrated inference. All work is verification-first and grounded in real execution traces and hardware behavior.
Infrastructure for proving model behavior through hardware-level instrumentation. Direct verification of what models compute at the instruction level, bypassing interpretability approximations.
Custom kernel and memory hierarchy designs targeting real hardware constraints. Maintaining behavioral fidelity while optimizing for latency and resource efficiency.
Understanding neural computation through compression. Compression guides both system design and interpretability analysis — what compresses well is what matters.
Direct interface to the lab. Type help for available commands.
RA1 Labs is the parent organization of Cybears — a competitive Valorant esports organization founded and led by Kanishk Joshi as IGL.
Competitive Valorant organization at Diamond level, building toward VCT participation. Cybears operates on the same principles as RA1's research: discipline, systematic improvement, verification-first decision making, and understanding before execution.
The organization runs competitive scrims, ranked improvement programs, and strategies informed by the same analytical rigor applied to research — the intersection of esports excellence and technical depth.
Independent researcher building AI systems from first principles. At 17, he has designed and implemented complete neural inference stacks, worked on mechanistic interpretability infrastructure, and contributed to hardware-software co-design research.
He operates without institutional affiliation, funding, or external partnerships — enabling research decisions driven by technical truth rather than institutional pressure.
His approach is systematic and verification-first: understand the problem deeply, build the solution carefully, verify the results rigorously. He believes intelligence — human and artificial — is not inherited but built, piece by piece, proof by proof.
RA1 operates as a solo research lab. No external funding. No committees. No compromises on research direction. This enables speed and intellectual honesty — we pursue what is true rather than what is fundable, accountable directly to the technical reality of what we build.
We work in layers: understanding before building, building before deploying, verifying everything before claiming results. Hardware efficiency is not optional — computational constraints force clear thinking. We design from scratch when existing tools do not match our requirements.
For research inquiries, partnerships & collaboration
Cybears inquiries — scrims, roster, partnerships — use the same contact.