ML Systems
cpptensor
CUDA Kernels, C++
Frontier AI research
Frontier AI research, enterprise systems, and scientific applications are expressions of a shared pursuit of understanding and building systems that operate reliably in complex, real-world environments, where each domain reinforces and advances the others. Research Commons works across it all.
Backed by
03+
We’re building on original, foundational research — every platform, product, and partnership we ship is powered by it. Enterprise systems, open exploration, and education feed one another — and each pass through that loop compounds into systems that hold up in the real world.
How we do it
Building end-to-end AI infrastructure engineered to operate at scale. From training and inference to agentic systems, and everything in between.
Enterprise
Kubernetes and Slurm-native infrastructure for reliable, low-config distributed training — deployable on your cloud, clusters, or enterprise compute.
Workload-aware inference optimization across vLLM, SGLang, LMDeploy, and TensorRT — reducing GPU idle time, serving cost, and latency through automatic workload simulation and bottleneck profiling across your infrastructure stack.
High-performance agent runtimes with optimized sandboxing, trajectory observability, and dedicated RL environments — built for secure, production-grade browser, API, and long-horizon agents.
Optimization across training and inference aimed at improving utilization, latency, and throughput under real workloads.
Keep weights, runtimes, and deployments in your infrastructure, with systems designed to remain inspectable and operable.
Architecture built to evolve with changing models, hardware, and operating demands without repeated re-architecture.
AionNeocloud
LanturnBehavioural Data
Pavo AIEnterprise Intelligence
Exploration
Open systems, benchmarks, and
tools across frontier domains
ML Systems
cpptensor
CUDA Kernels, C++
Medical AI
ASD-Bench
Cognitive Neuroscience
ML Systems
cppnet
Networking, RPC, C++
Medical AI
Spine-Bench
Image Segmentation
Medical AI
MedSAM-RL
Reinforcement Learning
Document AI
ParseTens
Deterministic PDF Parsing
Medical AI
Medical World Models
Model-Based RL
Scientific Computing
Bayesian Neural ODEs
SciML 25, Naples
ML Systems
cppgrad
NN Frameworks, C++
Medical AI
05 / 09
Publication
Under Review
Stage
Research in Progress · Medical Imaging
Methods
Reinforcement Learning, MedSAM, Medical Imaging, Segmentation
Cascading reinforcement learning for medical image segmentation.
Medical image segmentation is often treated as a one-shot prediction problem, despite clinical workflows requiring iterative refinement and anatomical reasoning. MedSAM-RL reframes segmentation as a long-horizon sequential decision process, where cascaded reinforcement learning agents refine masks step by step on top of MedSAM. The framework explores whether RL-driven refinement can improve robustness, edge-case handling, and label efficiency in medical imaging settings where expert annotations are scarce.
A cascaded reinforcement learning framework that iteratively refines MedSAM segmentation outputs through sequential decision-making, optimizing anatomical consistency and segmentation quality across clinical imaging tasks.
Explores a more adaptive approach to medical segmentation by combining foundation segmentation models with long-horizon RL, improving robustness in challenging clinical cases and reducing dependence on dense annotations.
Education
Spanning problem-solving, knowledge building and research
Learn by solving
Curated problem sets built on top of MIT OCW, designed for people who learn by solving, not watching. Concepts paired with structured practice, fast feedback, and tight progression loops.
VisitEnd-to-end research
A research platform for reading, writing, and working through ideas end-to-end. Papers, notes, citations, and analysis—kept in one continuous workflow instead of scattered tools.
Know more
Blog
Blog
Blog
Blog
Blog
Blog
Blog
