{
  "route": {
    "id": "ml-infrastructure",
    "label": "ML Infrastructure",
    "short": "MLI",
    "summary": "Retrieval, inference serving, and distributed training.",
    "projects": [
      "annlite",
      "vllm-lite",
      "dist-train"
    ]
  },
  "owner": {
    "name": "Aaryan Patel",
    "email": "aaryansp26@gmail.com",
    "github": "https://github.com/asp53826",
    "linkedin": "https://www.linkedin.com/in/aaryanpatelsystems/",
    "site": "https://asp53826.github.io",
    "role": "Systems and ML Infrastructure Engineer",
    "education": "Computer Science undergraduate at the University of Georgia",
    "work": "Automation and analytics at MP Equipment"
  },
  "thesis": "I build systems that survive faults, bad data, and dishonest benchmarks.",
  "projects": [
    {
      "id": "annlite",
      "name": "annlite",
      "track": "ml-infrastructure",
      "language": "C++17 + Python",
      "tagline": "HNSW with SIMD kernels, Python bindings, and FAISS comparison.",
      "problem": "Approximate retrieval claims are meaningless without the full recall-versus-latency curve.",
      "mechanism": "Layered HNSW graph, diversity-aware neighbor selection, SIMD distance kernels, and GIL-free search.",
      "attack": "Exact ground truth, recall-matched FAISS comparison, and thread-count invariance.",
      "metric": "1.83x",
      "metricLabel": "FAISS QPS at 0.999 recall",
      "proof": "51 tests; at 98.5% recall it performs 26x fewer distance computations than exact scan.",
      "limitation": "FAISS remains 10-20% faster in the low-recall regime; the crossover appears only at high recall.",
      "command": "git clone https://github.com/asp53826/annlite && cd annlite && pip install -r requirements-dev.txt && make test",
      "repo": "https://github.com/asp53826/annlite",
      "methods": [
        "exact recall",
        "FAISS baseline",
        "concurrency invariants"
      ]
    },
    {
      "id": "vllm-lite",
      "name": "vllm-lite",
      "track": "ml-infrastructure",
      "language": "Python",
      "tagline": "Paged KV cache and continuous batching around a small model.",
      "problem": "Static batching strands KV memory and creates head-of-line blocking under variable request lengths.",
      "mechanism": "Paged cache, continuous batching, chunked prefill, prefix caching, and speculative decoding.",
      "attack": "Same weights, sampling, traffic distribution, and memory budget across sequential, static, and paged schedulers.",
      "metric": "94%",
      "metricLabel": "KV slot utilization vs 21%",
      "proof": "90 tests; paged scheduling reaches 2.9x sequential throughput and 5.5x better median TTFT.",
      "limitation": "CPU-scale simulator with a small model; it does not claim production GPU kernel performance.",
      "command": "git clone https://github.com/asp53826/vllm-lite && cd vllm-lite && pip install -r requirements.txt && pytest",
      "repo": "https://github.com/asp53826/vllm-lite",
      "methods": [
        "scheduler benchmark",
        "memory accounting",
        "token invariance"
      ]
    },
    {
      "id": "dist-train",
      "name": "dist-train",
      "track": "ml-infrastructure",
      "language": "Python",
      "tagline": "Ring all-reduce built from send/recv, then measured where it loses.",
      "problem": "A one-line collective hides the communication cost and the scale at which latency dominates bandwidth efficiency.",
      "mechanism": "Ring all-reduce, data parallelism, tensor parallelism, pipeline parallelism, and explicit cost models.",
      "attack": "Every collective is checked against torch.distributed across worker counts and tensor sizes.",
      "metric": "58.7 MB",
      "metricLabel": "per worker at eight workers",
      "proof": "57 tests; the naive baseline sends 234.9 MB per worker under the same eight-worker case.",
      "limitation": "Loopback gloo on a 10-core machine; efficiency falls past four workers and small tensors favor naive collectives.",
      "command": "git clone https://github.com/asp53826/dist-train && cd dist-train && pip install -r requirements.txt && pytest",
      "repo": "https://github.com/asp53826/dist-train",
      "methods": [
        "torch.distributed oracle",
        "cost model",
        "scaling sweep"
      ]
    }
  ],
  "generated": "2026-07-31"
}