Linux Platform & AI Infrastructure Engineer
Focused on low-level Linux systems programming, deterministic C++17 memory models, kernel DRM/TTM subsystems, and resilient L7 AI gateways.
[ Physical AMD RDNA 2 GPU / Linux Kernel 6.x ]
│
▼ pread(fd, buf, 64, 0)
+─────────────────────────────────────+
| amdgpu-sentinel | • C++17, epoll_pwait2, timerfd, signalfd4
| (Daemon: 420 KB RSS, cgroups v2) | • Zero-alloc hot path, lockless VFS read
+──────────────────┬──────────────────+
│
▼ POSIX SHM (mmap, Seqlock <100 ns)
/dev/shm/vram_guard
▲
│ Zero-copy telemetry read
+──────────────────┴──────────────────+
| prompt-mesh | • Python 3.12 asyncio, RSS ~13 MB
| (L7 Gateway: OpenAI API :8000) | • Middle-Out Pruning (anti-TTM spillover)
+──────────────────┬──────────────────+ • Active cancellation on client abort
│
▼ Proxy to local cluster
llama-server (:8080 / :8082)
───────────────────────────────────────────────────────────────────────────
Testing & DevSecOps: [sentinel-mesh-ops]
• Headless Bubblewrap (bwrap) sysfs mocking in GitHub Actions (Ubuntu 24.04)
• Zero-Allocation proof via GNU ld linker wrapping: -Wl,--wrap=malloc
• Multi-compiler sanitizer matrix: ASan / UBSan / TSan (Seqlock verification)
mkdir -p ~/BannedForWhat
cd ~/BannedForWhat
git init
git branch -M main
cat << 'EOF' > README.md
# Dmitry Galoshin (@BannedForWhat)
**Linux Platform & AI Infrastructure Engineer**
Focused on low-level Linux systems programming, deterministic C++17 memory models, kernel DRM/TTM subsystems, and resilient L7 AI gateways.
---
### Architecture: Sentinel-Mesh Inference Infrastructure
```text
[ Physical AMD RDNA 2 GPU / Linux Kernel 6.x ]
│
▼ pread(fd, buf, 64, 0)
+─────────────────────────────────────+
| amdgpu-sentinel | • C++17, epoll_pwait2, timerfd, signalfd4
| (Daemon: 420 KB RSS, cgroups v2) | • Zero-alloc hot path, lockless VFS read
+──────────────────┬──────────────────+
│
▼ POSIX SHM (mmap, Seqlock <100 ns)
/dev/shm/vram_guard
▲
│ Zero-copy telemetry read
+──────────────────┴──────────────────+
| prompt-mesh | • Python 3.12 asyncio, RSS ~13 MB
| (L7 Gateway: OpenAI API :8000) | • Middle-Out Pruning (anti-TTM spillover)
+──────────────────┬──────────────────+ • Active cancellation on client abort
│
▼ Proxy to local cluster
llama-server (:8080 / :8082)
───────────────────────────────────────────────────────────────────────────
Testing & DevSecOps: [sentinel-mesh-ops]
• Headless Bubblewrap (bwrap) sysfs mocking in GitHub Actions (Ubuntu 24.04)
• Zero-Allocation proof via GNU ld linker wrapping: -Wl,--wrap=malloc
• Multi-compiler sanitizer matrix: ASan / UBSan / TSan (Seqlock verification)
- amdgpu-sentinel — Zero-allocation C++17 telemetry daemon for AMD RDNA 2 sysfs (420 KB RSS, systemd cgroups v2).
- prompt-mesh — Hardware-aware L7 AI Gateway on pure asyncio, middle-out context pruning, and circuit breaker.
- sentinel-mesh-ops — Automated reliability harness, bwrap sysfs emulation, linker allocation checks, and sanitizer matrices.
Linux 6.x • C++17 (no-RTTI/no-exceptions) • POSIX IPC • cgroups v2 • Seccomp BPF • ROCm/HIP • Python asyncio • DevSecOps / SCA