We’re the team behind LMCache, the open-source engine built to eliminate redundant context recomputation in LLM inference. Tensormesh turns that research into production infrastructure, deployed across real GPU fleets, not just benchmarks.
In this video, Tensormesh CEO and Co-Founder Junchen Jiang explains how model-level caching reduces repeated work, lowers AI costs, and creates a foundation for context-heavy production workloads.
Our values
AI should scale by reusing work, not repeating it. We help teams spend less on redundant context and more on building better products.
We validate before we ship. Every feature in Tensormesh Operator traces back to a measurable gain in LMCache.
Customers should know exactly what they pay for and what they save. We make caching, pricing and performance measurable.
Context-heavy AI is a systems problem. We design for the full path from request to cache to model response.
Our story
Every Tensormesh capability started as a measured result in LMCache, the open-source project our founders created.
KV caching research led by Junchen Jiang, Yihua Cheng and Kuntai Du.
More than 6,000 GitHub stars across the two projects.
Awarded for CacheBlend, non-prefix KV cache reuse.
Tensormesh launches to cut inference cost and latency by up to 10x.*
From investors including AMD Ventures, CoreWeave and NVentures.
The data management layer for AI inference, generally introduced.
*Results may vary based on specific client infrastructure, workload characteristics, and hardware configuration.
Meet the team
Junchen Jiang is CEO and Co-Founder of Tensormesh, and an Associate Professor of Computer Science at the University of Chicago. His research focuses on networked systems and their intersections with machine learning, with recent emphasis on making large language model (LLM) inference faster and more cost-efficient. He is a faculty open-source lead for the LMCache Lab, where he co-leads the development of LMCache, a state-of-the-art KV caching library, and the vLLM Production Stack for large-scale LLM deployment.
Junchen’s work has bridged academia and industry, from pioneering the first data-driven Internet-scale optimization systems for video streaming quality (Conviva) and Internet telephony quality (Skype) to building adaptive edge video analytics systems (Microsoft). Between 2022 and 2024, he also served as an advisor to Conviva, where he co-created the company’s new real-time data analytics system, with Yihua Cheng as the main contributor. He has collaborated broadly with industry leaders, including NVIDIA, Microsoft, and Conviva, to translate research breakthroughs into production systems.
He received his bachelor’s degree from Tsinghua University’s Yao Class in 2011 and his Ph.D. in Computer Science from Carnegie Mellon University in 2017, where he was awarded the School of Computer Science Distinguished Dissertation Award. His research has been recognized with an NSF CAREER Award, two Google Faculty Research Awards for large-scale AI systems, and a Best Paper Award at EuroSys for his work on KV caching systems.
Yihua Cheng is Co-Founder and CTO of Tensormesh, and a computer systems and AI infrastructure expert with deep expertise in large language models, high-performance computing, and open-source development. He earned his Ph.D. in Computer Science from the University of Chicago and his bachelor’s degree from Peking University.
During his Ph.D., Yihua created LMCache and the vLLM production stack, open-source projects that have collectively earned more than 6,000 GitHub stars. His research spans computer systems, machine learning, and large-scale LLM deployment, and he brings end-to-end expertise from low-level hardware to state-of-the-art ML algorithms. He is a co-recipient of the ACM EuroSys 2025 Best Paper Award for CacheBlend.
Yihua also has significant industry experience. At Conviva, Inc., he spent three years as a research engineer, where he single-handedly built the company’s next-generation real-time data analytics system. He has also interned at Microsoft Research, developing infrastructure for real-time video systems, and at Alibaba and Megvii, contributing to production development across the stack from hardware drivers to user-facing applications.
Earlier in his career, Yihua was part of Peking University’s High Performance Computing Team, where he competed in international HPC competitions and won multiple awards. As a high school student, he ranked among the top 50 in China’s Chemistry Olympiad, an early reflection of his strong problem-solving and analytical skills.
Kuntai Du is the Chief Scientist and Co-Founder of Tensormesh. He completed his Ph.D. in Computer Science at the University of Chicago in 2025, where he focused on bridging cutting-edge research with real-world LLM infrastructure. His work centers on LLM inference engines and knowledge management, and he has authored more than 15 papers at top systems and machine learning venues, including SIGCOMM, NSDI, OSDI, SOSP, SoCC, EuroSys, and ASPLOS. His contributions have been recognized with the Siebel Scholar award for outstanding academic achievement and leadership. He has also driven the landing of novel research techniques into open-source projects such as vLLM, LMCache, and the vLLM Production Stack.
Before co-founding Tensormesh, Kuntai was a core contributor to the vLLM project, where he built performance benchmarking infrastructure for regression tracking and cross-engine comparisons in collaboration with the NVIDIA TensorRT-LLM team. He also designed and opened the KV cache connector interface, enabling contributions from multiple companies, including RedHat’s LLM-D team, Moonshot AI, and Alibaba, while allowing organizations such as Meta and NVIDIA to extend the API for internal use. This interface has since been adopted by the NVIDIA Dynamo project.
Beyond research, Kuntai has industry experience at Microsoft Research, where he worked on large-scale imagery compression, and at TuSimple, where he developed video understanding pipelines for autonomous driving.
Chief Product Officer
Sandro Mazziotta, PhD, is Chief Product Officer at Tensormesh, an innovative startup focused on optimizing AI inference and KVCache efficiency. He is a catalyst for growth, specializing in partnering with executive teams to build and scale transformative technology businesses by translating emerging technologies into significant, revenue-generating market opportunities.
His extensive career as an “intrapreneur” at major tech leaders includes building the Red Hat OpenStack Telco business from inception to over $100M and pioneering the company’s global sustainability strategy. He has also held key roles at HPE (Hewlett Packard Enterprise).
His advisory approach is uniquely hands-on, bridging boardroom strategy and engineering reality. Dr. Mazziotta holds a Computer Science PhD from Nice University and leads elite technical teams, providing the strategic guidance and entrepreneurial drive needed to find product-market fit, achieve exponential growth, and build lasting global value.
Head of Engineering
Nick Barcet is Vice President of Engineering at Tensormesh, where he drives efficiency and stability across all engineering efforts in coordination with the vision set by the company’s CEO and CTO. Prior to Tensormesh, Nick was Head of the Field CTO organization at Red Hat. He joined Red Hat following its acquisition of eNovance, a startup where he served as Head of Products and Presales. He has more than 30 years of experience working for hardware and software vendors and startups alike, and has focused on open source development models since the early 2000s.
Head of Go-To-Market
Nigel Thomas is Head of Go-To-Market at Tensormesh, leading sales, partnerships, marketing, and customer success for the company’s AI inference optimization platform. Nigel brings more than two decades of infrastructure go-to-market leadership and a maniacal focus on customer satisfaction that builds lasting, trusted relationships.
He has led sales at four companies through successful acquisitions: FuseSource and Inktank, both acquired by Red Hat (Inktank for $175M, becoming a core part of Red Hat’s Emerging Technology strategy); npm, acquired by GitHub; and most recently Moesif, acquired by WSO2, where he served as Head of Sales. At Red Hat, he was also part of the Emerging Technologies leadership team driving net-new business through the company’s $34B acquisition by IBM in 2019.
Known for building worldwide sales teams from the ground up, closing marquee early deals for startups, and a process-driven, KPI-focused approach, Nigel has also spoken at conferences worldwide on storage, integration technology, and open source. He holds a degree in Electrical and Electronic Engineering from Trinity College Dublin.
We’re hiring engineers and researchers to solve the industry’s biggest efficiency challenges.
