AI infrastructure lab
Building the infrastructure
for modern AI.
Making training and inference easier for the world.
Research Commons is an AI infrastructure lab building the systems required to train, serve and continuously improve AI models.
We believe modern AI is not only a machine-learning problem. It is also a distributed-systems and platform-engineering problem.
Improving a model requires more than choosing the right algorithm. Teams must coordinate environments, rollouts, inference workers, reward models, verifiers, distributed training jobs, evaluations, checkpoints and production deployments across increasingly heterogeneous compute.
Today, every serious AI team is forced to build this infrastructure for itself. Researchers spend time operating clusters, diagnosing failed jobs, moving checkpoints and connecting fragmented systems instead of improving model capabilities.
We are building two layers to change that.
01 · Tensile
The infrastructure plane
Tensile is the infrastructure layer for agentic companies and neolabs. It runs first-party on compute you already own, clusters, clouds, neoclouds or on-prem, and ships as three products: Tensile-Train, Tensile-Infer and Tensile-Agents.
Tensile-Train runs distributed post-training end to end, SFT, DPO, ORPO, PPO and GRPO++, K8s- and Slurm-native. Tensile-Infer runs workload-aware distributed inference across vLLM, SGLang, LMDeploy and TensorRT. Tensile-Agents is the runtime for production agents: sandboxes, RL environments, rollouts and evaluations.
One infrastructure plane across every cloud. Run the same train, infer and agent stack on any public cloud, neocloud or on-prem cluster, without rewrites or vendor lock-in.
As compute grows more heterogeneous, teams should be free to run wherever capacity is cheapest and closest, not rebuild their stack for each provider. Tensile gives teams a reliable, portable foundation to train, serve and run agents on hardware they control.
# one plane: train · infer · agents$ tensile launch train.yaml --cloud any ✓ scheduled · neocloud a100×8 · us-west · $1.91/hr$ tensile infer serve --engine vllm ✓ workload-aware · same cluster, same plane$ tensile run -- python rollout.py ✓ sandbox · k8s, slurm, on-prem: same command02 · Niko
The autonomous post-training engineer
On top of Tensile, we are building Niko, an autonomous engineer that understands the complete post-training system.
Niko helps teams configure workloads, deploy infrastructure, diagnose failures, optimize GPU utilization, recover experiments and find the fastest path from a base model to a production-qualified model.
Over time, Niko will learn from every workload it operates, creating a systems model of how models, algorithms, environments and distributed infrastructure interact.
Tensile gives Niko a system to operate. Niko gives Tensile the intelligence to improve it.
Our aim is to eliminate AI infra and platform engineering as a job role so teams can focus on what matters most to them, without compromising on performance and quality.
We believe everyone has the right to be frontier. Just as coding had its disruptive moment, Niko will bring that shift to post-training.
$ niko "my SFT run OOMs at step 4k" → reading logs, gpu memory, config cause: grad checkpointing off, seq_len 8192 fix: enabled checkpointing, mbs 4 to 2 · re-queued$ niko optimize --goal "cut cost 30%, keep eval" ✓ moved to spot neocloud · same eval · -34% $/runWho we are
A research lab, building in the open.
Research Commons is a research-driven infrastructure company, a team of researchers and systems engineers building the infrastructure layer for modern AI. We conduct frontier research and ship the systems it runs on, Tensile and Niko, sharing our work through the Research Commons Lab.
- Active research across programmatic PDF parsing, medical world models and RL-based segmentation.
- Tensile sandboxes, our flagship: state-of-the-art, secure and reproducible environments for running rollouts, evaluations and agentic workloads across any cluster.
- Tensile-Train and Tensile-Infer: distributed post-training and workload-aware inference on your own compute, benchmarked against Kubeflow, KubeRay, SkyPilot and Volcano.
- Public SDKs with 10K+ downloads, plus open-source C++ tensor and autograd libraries.
- RL environments and world models delivered to frontier labs.
Team and advisors from Google, Microsoft, OpenAI, DeepMind, Mercor and YC-backed startups.
Research published at ICML, EMNLP, MSML and CVIP.
Our goal is simple
AI teams should focus on improving intelligence. The infrastructure required to produce it should operate itself.
Explore our research