RightNow Research Lab
Enabling Model-Hardware Co-Design at Scale
Abstract
We build the tools that close the gap between AI models and the hardware they run on, giving engineers and enterprises the infrastructure to go from model to optimized production inference, faster.
NVIDIA Inception member · Backed by Y Combinator · 6 papers on arXiv
Products
We maintain infrastructure and developer tools for GPU programming, inference, and deployment.
RunInfra
NewGPU infrastructure for deploying, measuring, and operating inference workloads.
Research
Work on sparse attention, dynamic model adaptation, world models, token-level inference, and automated GPU-kernel search. Papers are available on arXiv.
AutoMegaKernel: A Statically-Checked Agent Harness for Self-Retargeting Megakernel Synthesis
AutoKernel: Autonomous GPU Kernel Optimization via Iterative Agent-Driven Search
StreamIndex: Memory-Bounded Compressed Sparse Attention via Streaming Top-k
Ouroboros: Dynamic Weight Generation for Recursive Transformers via Input-Conditioned LoRA Modulation
HCLSM: Hierarchical Causal Latent State Machines for Object-Centric World Modeling
TIDE: Token-Informed Depth Execution for Per-Token Early Exit in LLM Inference
Open Source
We maintain open-source systems for agent infrastructure, edge inference, and automated kernel optimization.
OpenFang
Rust system for low-level agent execution with direct access to operating-system and GPU interfaces.
Inspect OpenFang star samples
| Sample | Stars |
|---|---|
| 1 | 1 |
| 2 | 1,591 |
| 3 | 3,211 |
| 4 | 4,831 |
| 5 | 6,421 |
| 6 | 8,041 |
| 7 | 9,661 |
| 8 | 11,281 |
| 9 | 12,871 |
| 10 | 14,491 |
| 11 | 15,600 |
| 12 | 17,864 |
PicoLM
Small C inference runtime for memory-constrained edge devices.
Inspect PicoLM star samples
| Sample | Stars |
|---|---|
| 1 | 1 |
| 2 | 151 |
| 3 | 331 |
| 4 | 511 |
| 5 | 691 |
| 6 | 871 |
| 7 | 1,081 |
| 8 | 1,261 |
| 9 | 1,350 |
| 10 | 1,410 |
| 11 | 1,445 |
| 12 | 1,656 |
AutoKernel
Kernel-search system that profiles PyTorch models, tests Triton variants, and uses measured bottlenecks to choose follow-up experiments.
Inspect AutoKernel star samples
| Sample | Stars |
|---|---|
| 1 | 1 |
| 2 | 91 |
| 3 | 211 |
| 4 | 331 |
| 5 | 451 |
| 6 | 571 |
| 7 | 691 |
| 8 | 811 |
| 9 | 860 |
| 10 | 890 |
| 11 | 905 |
| 12 | 1,418 |
