Skip to content
View BannedForWhat's full-sized avatar

Block or report BannedForWhat

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
BannedForWhat/README.md

Dmitry Galoshin (@BannedForWhat)

Linux Platform & AI Infrastructure Engineer | Systems DevSecOps
Focused on low-level Linux systems programming (C++17/C++20 bare-metal), deterministic zero-allocation memory models, kernel DRM/TTM telemetry, unprivileged containerless runtime sandboxing, and resilient L7 AI gateways.


Architecture: Sentinel-Mesh & Agent-Jail Infrastructure

[ Hardware: AMD RDNA 4 (Navi 48 XTX, gfx1201) / Linux Kernel 6.x / Arch Linux ]
                                │
                                ▼ persistent FDs, pread(fd, buf, 64, 0)
┌─────────────────────────────────────────────────────────────────────────────┐
│ 1. amdgpu-sentinel (L1/L2 Kernel Telemetry Daemon)                          │
│    • Bare-metal C++17 (-fno-exceptions, -fno-rtti, -O2)                     │
│    • Reactor: epoll_pwait2 + timerfd (CLOCK_MONOTONIC) + signalfd4          │
│    • Footprint: 420 KB RSS, 0 heap allocations in hot loop                  │
└──────────────────────────────────────┬──────────────────────────────────────┘
                                       ▼ POSIX SHM (mmap, Seqlock <50 ns)
                               /dev/shm/vram_guard
                                       │
                                       ▼ Zero-copy atomic read
┌─────────────────────────────────────────────────────────────────────────────┐
│ 2. prompt-mesh (L7 VRAM-Aware AI Gateway)                                   │
│    • Pure Python 3.12 asyncio (Port :8000, zero third-party dependencies)   │
│    • Anti-OOM: Middle-Out Pruning triggered by VRAM >85% or ttm_alert == 1  │
│    • Active Slot Cancellation: immediate TCP reset + POST /slots/erase      │
│    • Recency Bias Anchors & Intent Router (:8080 Code, :8082 Architecture)  │
└──────────────────────────────────────┬──────────────────────────────────────┘
                                       │
            ┌──────────────────────────┴──────────────────────────┐
            ▼ HTTP Stream                                         ▼ Pathname UNIX Socket
┌───────────────────────────────┐             ┌───────────────────────────────────────────────┐
│ llama-server (:8080 / :8082)  │             │ 3. agent-jail (Runtime Security Sandbox)      │
│ • ROCm 7.x / Mesa RADV Vulkan │             │    • C++20 containerless isolator (<1 ms cold)│
│ • Qwen2.5-Coder-14B (Q5_K_M)  │             │    • 6 Namespaces (user, mount, pid, net, ...)│
│ • 100% VRAM, 32K context q8_0 │             │    • Arch Linux Merged /usr VFS + pivot_root  │
└───────────────────────────────┘             │    • Landlock LSM (ABI v5) + Seccomp-BPF      │
                                              │    • cgroups v2: memory.swap.max=0, pids.max  │
                                              │    • Non-blocking PTY master/slave forwarder  │
                                              └───────────────────────────────────────────────┘

───────────────────────────────────────────────────────────────────────────────────────────────
4. sentinel-mesh-ops (DevSecOps & Reliability Verification Harness)
   • Headless Bubblewrap (bwrap) sysfs mocking in GitHub Actions (Ubuntu 24.04 under AppArmor)
   • Linker-level proof of zero allocations via GNU ld: -Wl,--wrap=malloc/calloc/realloc/free
   • Isolated multi-compiler sanitizer matrix: ASan / UBSan / TSan (Seqlock race verification)

Verified Production Repositories

  • agent-jail — Sub-millisecond (<1 ms) unprivileged containerless runtime sandbox for autonomous AI coding agents in C++20. Implements 6 Linux namespaces, Arch Merged /usr VFS with pivot_root(2), Landlock LSM (ABI v5), handcrafted Seccomp-BPF filters (clone3 -> ENOSYS, AF_UNIX enforcement), direct cgroups v2 resource capping (memory.swap.max=0), and asynchronous PTY multiplexing.
  • amdgpu-sentinel — Zero-allocation C++17 bare-metal telemetry daemon for AMD DRM/TTM sysfs. Uses lockless pread() at offset 0 to bypass VFS f_pos_lock, single-threaded epoll_pwait2 reactor with monotonic timerfd, and atomic 64-byte Seqlock state export to /dev/shm/vram_guard (420 KB RSS, systemd user service sandbox).
  • prompt-mesh — Hardware-aware, high-performance L7 AI Gateway on pure Python 3.12 asyncio. Reads kernel telemetry via zero-copy mmap struct unpacking (<100 ns), executes microsecond Middle-Out Pruning on VRAM pressure, injects Recency Bias Anchors, and terminates backend decoder loops via active socket resets on client disconnect.
  • sentinel-mesh-ops — Comprehensive CI/CD reliability harness. Features headless sysfs tree mocking via rootless Bubblewrap with dynamic AppArmor profiles, binary zero-allocation proof via GNU ld symbol wrapping, and race-condition validation under ThreadSanitizer (TSan).

Core Engineering Invariants

  • Linux Kernel Internals: Namespaces (CLONE_NEWUSER, CLONE_NEWNS, CLONE_NEWPID, CLONE_NEWNET), Landlock LSM (ABI v1–v5), Seccomp-BPF, cgroups v2 (memory.max, memory.swap.max=0, pids.max), VFS pivot_root(2), epoll_pwait2, timerfd, signalfd4, PTY line discipline.
  • Low-Level Systems (C++17 / C++20): Bare-metal systems programming (-fno-exceptions, -fno-rtti, -O2), lockless Sequence Locks (Seqlock), persistent file descriptors, zero heap allocations in execution hot loops, POSIX shared memory (/dev/shm).
  • AI Infrastructure & Acceleration: Local LLM deployment on AMD RDNA 4 (gfx1201), ROCm/HIP and Vulkan RADV cooperative matrices, FlashAttention-2 buffer optimization, quantized KV caching (q8_0), prompt routing, slot lifecycle management.
  • DevSecOps & Software Supply Chain: Rootless process sandboxing, linker interception (-Wl,--wrap), Clang Sanitizers (ASan, UBSan, TSan), SBOM generation (CycloneDX 1.6 / SPDX 3.0), SCA/OSA dependency quarantine, and mitigation of dependency confusion attacks.

Contact & Artifacts

Popular repositories Loading

  1. amdgpu-sentinel amdgpu-sentinel Public

    Zero-allocation C++17 telemetry daemon for AMD RDNA 2 GPUs. Reads sysfs via pread(), runs on epoll_pwait2/timerfd, exports state via POSIX SHM Seqlock, isolated in systemd user session (420 KB RSS).

    C++

  2. prompt-mesh prompt-mesh Public

    Lightweight Python asyncio AI gateway. Reads GPU telemetry via mmap from POSIX SHM, performs middle-out context pruning to prevent VRAM eviction, and handles active slot cancellation for llama.cpp.

    Python

  3. sentinel-mesh-ops sentinel-mesh-ops Public

    CI/CD reliability harness for amdgpu-sentinel and prompt-mesh. Features headless Bubblewrap sysfs emulation on Ubuntu 24.04, linker-wrapped zero-allocation tests, and ASan/TSan matrices.

    C++

  4. BannedForWhat BannedForWhat Public

  5. agent-jail agent-jail Public

    Sub-millisecond (<1ms) unprivileged containerless runtime sandbox for autonomous AI coding agents in C++20 (Namespaces, Landlock LSM, Seccomp-BPF, cgroups v2, PTY).

    C++