Nick Blokhin is a software architect and engineering team lead at GoBolt. His independent work includes GridCore, a scheduler that coordinates several local language models sharing one GPU.[1]
Start with GridCore – a scheduler that lets several local LLMs share one GPU →
Links
Personal biography
GitHub · w512 Software Engineering Team Lead · GoBolt
Career
Blokhin’s CV records leading software engineering at GoBolt from April 2025, following work as Lead AI Engineer and Software Architect at Instock from May 2021 to April 2025. Earlier he was a principal engineer and then technical team lead at Takeoff Technologies. His responsibilities combine application architecture, warehouse and commerce systems, and managing engineering teams.[2]
Contributions
GridCore sits between applications and llama.cpp server processes. It tracks model memory, residency and queued requests, then decides which work can run on the GPU. Interactive chat can receive priority over document indexing, while eviction and admission decisions prevent several independent applications from exhausting memory. Blokhin’s article explains adjustments needed after testing on an actual 16GB card, separating scheduling policy from the inference engine that executes a model.[3]
Tools, writing & talks
Writing, research & talks
Start here
An implementation account of request priorities, memory accounting and model residency for a shared GPU.
Read the control-plane diagram and one failure from the real-hardware section. Experience with local model serving helps; the reported configuration uses llama.cpp processes and a constrained GPU, so reproduce its assumptions before generalizing.
Author and GridCore developer[3]Original source