Home
Posts
-
Releasing the GPU while your agent waits is the obvious move. It saves nothing.
-
KV-cache reuse is an energy lever. Per-token attribution can't see it.
-
CUDA graphs always speed the kernel. They don't always speed the server.
-
Four GPUs, two sockets, one workload that didn't need any of it.
-
The client measured the cost. Only the per-device view measured the trade-off.
-
Disaggregation isn't a deployment topology. In llm-d, it's a per-request decision.
-
NVIDIA's KV-router isn't faster. Under load it drops requests — and that's the design.
-
I built an eBPF profiler to find where vLLM cold-start spends its time. It wasn't the disk.
-
The profiler had to teach me about the hardware. The hardware taught me about the profiler.
-
Profiling LLM inference: what your /proc sampler isn't telling you
-
Why your OpenTelemetry trace shows nothing useful when the CPU is doing all the work — a CP-SAT case study
-
How fragile is your weekly plan? A risk-premium framework for mid-market manufacturers
-
What an OR-Tools solver finds in a week of contract packaging — and what the planner usually misses
-
Three Production Scheduling Failures I've Seen, and the Math That Would Have Caught Them
-
Why your AI assistant can't actually plan your factory
-
Exposing a math solver as Circle Nanopayments: what I learned forking arc-nanopayments
-
How I exposed OR-Tools as a production MCP server
subscribe via RSS