FORWARD / WHY RADIXARK
Why RadixArk
A real serving foundation, a path into post-training, and an early systems problem worth shaping.
PUBLIC FOUNDATIONSINFERENCESGLangOpen, high-performance serving with real technical and community depth.
POST-TRAININGMilesRL and post-training on the other side of the improvement loop.
COMPANY STAGEEarly · open · infrastructure-firstFoundational interfaces and operating systems are still shapeable.
WHAT CONNECTS THEM→OBSERVABLE, RELIABLE
PRODUCTION SYSTEMS←
WHAT I BRINGSYSTEM DESIGNStable abstraction boundariesComplexity belongs below contracts that can survive scale and change.
OPERATIONSCross-layer context and control planesIdentity, topology, signals, changes, owners, diagnosis, and safe action.
EARLY-STAGE EXECUTION0→1 ownership with systems tasteMy edge is choosing what should remain durable—not typing faster than an agent.
TECHNICAL THESIS I WANT TO TESTAs agent context grows, KV and model-native state may become a first-class data layer.NOW · reuse + tiered placementNEXT · composition + intelligent servicesRESEARCH · state programming
A research direction—not a claim about RadixArk’s current roadmap.
RadixArk sits at the intersection of the future I want to build: open serving, post-training, and the systems that connect them.
Why RadixArkThat filter leaves very few companies. RadixArk matches it unusually well. The company already has real public foundations on both sides of the loop. SGLang is an open inference engine with genuine technical and community depth. Miles extends the foundation into reinforcement learning and post-training. RadixArk has also described an infrastructure-first mission focused on performance, correctness, reliability, hardware efficiency, managed infrastructure, and long-term system design.
I am specifically looking for an early company because the important interfaces are still shapeable. I do not claim that my advantage is writing code faster than an AI coding agent. My advantage is systems taste: deciding which abstraction should remain stable, where complexity belongs, what must be observable, and how an operating system can stay useful for years instead of one demo.
My past work maps naturally into that gap. I have built zero-to-one data platforms, normalized heterogeneous infrastructure, joined signals across layers and teams, encoded deterministic remediation, and then exposed bounded capabilities to multi-user agents. In an inference platform, the entities change—models, engine versions, GPUs, cache state, request cohorts, deployments, and evaluation cases—but the engineering discipline is similar.
I am also interested in the memory boundary. Today, KV reuse and placement are concrete serving problems. The next layer may treat model-native state as an intelligent data service that can be stored, scheduled, composed, and shared. Direct state programming is still research, and I would not present it as a mature product or as RadixArk’s roadmap. But it sits exactly at the intersection of serving systems, training, storage, and model behavior that I want to understand deeply.
My impression from the team and the public work is that RadixArk is early, technically ambitious, and serious about infrastructure as a creative discipline. It sits at the center of the future I want to build: open serving, post-training, and the observable, reliable systems that connect them.
Sources · RadixArk, “Our $100M Seed to Build Open Infrastructure for Frontier AI,” May 2026; SGLang and Miles public project materials; TensorMesh / LMCache interview. The KV extension and fit assessment are my own thesis, not a claim about RadixArk’s internal roadmap.