Bhargav bharqav

/* engine.c - baremetal compute & runtime descriptor */
#include <stdint.h>
#include <immintrin.h>
#include <sys/mman.h>

typedef struct __attribute__((aligned(64))) {
    const char *identity;           /* "bhargav // systems & infrastructure" */
    const char *target_arch;        /* "x86_64 [avx-512 vnni] + aarch64 [neon]" */
    
    struct {
        uint32_t zero_copy_hugepages : 1;  /* mmap(MAP_SHARED | MAP_HUGETLB) */
        uint32_t fused_int4_gemv     : 1;  /* _mm512_dpbusd_epi32 tensor compute */
        uint32_t lockfree_ring_buf   : 1;  /* single-producer multi-consumer IPC */
        uint32_t raft_wal_sync       : 1;  /* O_DIRECT zero-amplification append */
        uint32_t reserved            : 28;
    } hw_caps;

    const char *active_pipeline[3];
    uint64_t    cache_miss_budget;         /* 0x0ULL - non-negotiable */
} compute_engine_t;

static const compute_engine_t host = {
    .identity    = "Bhargav -> Systems & AI Infrastructure Engineer",
    .target_arch = "x86_64 (AVX-512 VNNI) / aarch64 (NEON)",
    .hw_caps = {
        .zero_copy_hugepages = 1,
        .fused_int4_gemv     = 1,
        .lockfree_ring_buf   = 1,
        .raft_wal_sync       = 1,
    },
    .active_pipeline = {
        "Sub-200MB 30B MoE inference via memory-mapped KV paging",
        "Lock-free ring buffers & actor IPC over POSIX shared memory",
        "VirtIO interrupt auditing & custom kernel dispatch routines"
    },
    .cache_miss_budget = 0ULL
};

Open Source Contributions

Crucible Security - Tool Injection Assessment Module
PR #65 · Issue #49

Built an adversarial attack engine covering OWASP AGENT-004 across MCP and tool-augmented agents. Four attack classes, twenty adversarial vectors, 286+ passing tests with dynamic attack registration.

Python Security MCP

Crucible Security - CI/CD Security Gating
PR #64 · Issue #52

Built a --fail-on severity threshold flag that blocks CI pipelines on HIGH/CRITICAL findings. Shipped reusable GitHub Actions templates for automated agent vulnerability scanning.

CLI CI/CD GitHub Actions

Microsoft OpenVMM - VirtIO Interrupt Fix
PR #4226

Found and fixed a spurious config-change interrupt during the DRIVER_OK transition. Audited INTx/MSI-X/MMIO interrupt paths across transports and corrected config_generation increment behavior.

Rust VirtIO Virtualization

youki (OCI Runtime) - Live Memory & cgroups v2
PR #3688

Fixed CLI argument propagation for --memory, --memory-reservation, and --memory-swap into the kernel cgroup layer. Corrected types to signed Option<i64> for unlimited allocations, added regression coverage.

Rust cgroups OCI

NVIDIA NodeWright - Lifecycle Drain Observability
PR #582 · Issue #542

Surfaced Blocked status conditions and transition-guarded warning events when non-interruptible workloads hold pre-drain barriers. Resolved condition-flapping bugs across reconciles and documented barrier semantics.

Go Kubernetes Operators

PyTorch ExecuTorch - Runtime Metadata Segfault Fix
PR #22405 · Issue #22404

Fixed a null pointer dereference in MethodMeta::uses_backend() when schema-optional FlatBuffer delegates are unset. Restored CMake fixture generation, re-enabled upstream method_meta_test, and synced Buck targets.

C++ FlatBuffers Runtime

Featured Systems

⚡ quantr-in-c

Zero-dependency quantized LLM/MoE inference in portable C99.

Fused SIMD GEMV using _mm256_maddubs_epi16 and _mm512_dpbusd_epi32
15 bit-exact test gates, paged Q8_0 KV cache
Sub-200MB peak RSS on 30B models

🐚 mysh

Production-grade POSIX mini-shell in C++.

Recursive AST parser for pipelines, subshells, redirection
Real job control with foreground/background process groups
Native directory stack, alias substitution, glob expansion

📦 distributed-kv-store

Fault-tolerant distributed KV store from first principles.

Raft consensus: leader election, log replication, snapshotting
Consistent hashing with virtual tokens
Hybrid logical clocks, append-only WAL

🔍 ultimate-hybrid-rag

High-throughput hybrid vector and lexical retrieval engine.

RRF fusion of dense embeddings and BM25 sparse indexes
Cross-encoder neural reranking pipeline
Sub-10ms latency on concurrent semantic chunking

🎬 tiny-video-diffusion

Efficient text-to-video latent diffusion built from scratch.

VideoDiT + VideoVAE with zero-init temporal attention
DPM-Solver++ scheduler and standalone C++ DDIM runtime
Faster-than-real-time CPU generation (RTF 0.17 on 8-frame clips)

📡 oemn

Offline Emergency Mesh Network in C11 over UDP.

Multi-hop Dijkstra shortest-path routing with binary min-heap
AES-256-GCM / ChaCha20-Poly1305 and sliding-window replay protection
Graceful degradation under 40% loss, sub-4s network reconvergence

🎨 tiny-diffusion

CPU-native latent diffusion system built from scratch.

Compact DiT and VAE architectures with Classifier-Free Guidance
Zero-dependency standalone C++17 inference runtime
1,000-image procedural verification gallery with full reproducibility

Technical Arsenal

Core Languages

Systems & Compute

Infrastructure & Data

Toolchain & Build

Currently Deep-Diving Into

Speculative decoding acceptance proofs - optimizing rejection sampling across batched draft verification steps
Zero-copy memory-mapped weight paging - eliminating page-fault penalties on sparse 30B MoE models under DRAM constraints
Lock-free ring buffers and actor models - low-latency IPC over POSIX shared memory

Navigation Menu