Go systems engineer, freelance. Nine years building production software — TypeScript, and in recent years Go, including contributions to ownCloud. Based in Czechia (CET).

Building now: memauditd — an open-source, read-only Linux agent that answers a question most fleets can't: how much of your DRAM is actually doing nothing?

It measures stranded host memory and KV-cache recompute waste on AI inference fleets using only what the kernel already provides — DAMON cold-page histograms in physical-address mode, PSI, cgroup v2 working-set counters, NVML, and vLLM's Prometheus metrics. No eBPF, no kernel modules, nothing to trust blindly: the whole agent is auditable in an afternoon.

$ memauditd selftest
kernel 6.8.0-45-generic ......... OK
cgroup v2 unified ............... OK
DAMON sysfs ..................... OK
DAMON tried_regions (>=6.2) ..... OK  → mode: full histogram
NVML ............................ OK (driver 560.x, 8 GPUs)
verdict: sampling mode available

Why now: the AI buildout made DRAM scarce and expensive, while the tooling to distinguish allocated from actually touched memory barely exists outside hyperscalers. Meta measured 20–30% cold memory at fleet scale. Almost nobody else has looked.

If you run GPU inference or a Linux fleet of ~25 hosts or more: I'm doing research interviews before locking the v1 design — twenty minutes of your war stories, or six questions answered in writing, whichever you prefer. Not a sales call. Reach me at ⟨lukas@hirt.cz⟩ or LinkedIn.

Elsewhere: ownCloud work · LinkedIn Freelance availability: occasionally, for Go and Linux systems work — ⟨lukas@hirt.cz⟩