Work

Selected projects

Each project is framed as a short case study: the problem, the approach, and the measurable impact.

Selected work

Projects that show how I think and ship

Each project is framed as a short case study: the problem, the approach, and the impact. If a metric isn’t available, I’m explicit about what changed.

Showing 6 items · Filter: All
  • Preview for graph-rca

    graph-rca

    Log-to-resolution RAG system over 200 annotated production incidents — LLM parses raw logs into a causal DAG stored in MongoDB; ChromaDB (HNSW index) retrieves runbook context scoped to that graph traversal before generating a root-cause report. Output quality validated by a 3-model judge ensemble (Qwen3:32b, GPT-4o-mini, Llama-3.1-70B) across 9 reproducible experiment scripts; ships with CPU/GPU Docker Compose variants and one-command ./run.sh setup.

    Software
    Problem
    Manual analysis of production incident logs for root cause determination is time-consuming and inconsistent.
    Tech
    PythonFastAPILangChainChromaDBMongoDB
    Impact
    Automated root-cause reports with high-quality validation and easy deployment, reducing resolution time.
    Source
  • Preview for permission-mongo

    permission-mongo

    RBAC-powered BaaS for MongoDB targeting 50K+ QPS custom AST-caching policy expression compiler, lock-free router with atomics, fasthttp tuned for 256K concurrent connections, batched async audit logging. Ships with labeled Prometheus histograms (latency, RBAC eval outcomes, cache hit/miss), a pre-built Grafana dashboard, and a Python SDK bundled alongside the Go service.

    Software
    Problem
    High-performance, scalable RBAC for MongoDB requires complex custom implementations.
    Tech
    GoMongoDBPrometheusGrafana
    Impact
    Delivers a production-ready RBAC service with advanced features, monitoring, and SDK support.
    Source
  • Preview for BetterBingImageDownloader

    BetterBingImageDownloader

    Published and actively maintained library with 18K+ downloads, 41 stars, 10 versioned releases; v3.0.0 was a correctness release fixing a thread-unsafe counter, atomic file writes, MD5 deduplication, and a global socket.setdefaulttimeout() side-effect. Test suite grew from 0 → 41 passing tests with GitHub Actions CI; follows full pip packaging lifecycle with pyproject.toml and entry-point CLI.

    Software
    Problem
    Existing tools for downloading Bing images were unreliable, lacking proper testing and packaging.
    Tech
    PythonPyPI
    Impact
    Reliable, well-tested library with significant adoption, improving dataset creation workflows.
    Live demoSource
  • Preview for Cyclomatic Complexity Clang Plugin

    Cyclomatic Complexity Clang Plugin

    Clang compiler plugin using AST traversal to compute per-function cyclomatic complexity across C/C++ codebases; generates refactoring reports and integrates with CI/CD to enforce complexity thresholds at review time. Awarded Best Project by HPE LLVM team; built during LLVM 2024 workshop alongside contributions to LLVM project toolchain.

    Software
    Problem
    Assessing and enforcing code complexity in large C/C++ projects is challenging without integrated tools.
    Tech
    C++Clang/LLVM
    Impact
    Automated complexity analysis tool that aids refactoring and code quality enforcement, recognized with an award.
  • Preview for Lung Classification

    Lung Classification

    Noise filter effects on deep learning medical imaging.

    Publication
    Problem
    Noise in CT and X-ray images impacts the accuracy of lung disease classification models.
    Tech
    PythonPyTorchImage Processing
    Impact
    Evaluated filter effects to improve model performance in medical imaging, published in IJEECS.
    Read paperSource
  • Preview for Handwritten Tripsheet Analysis

    Handwritten Tripsheet Analysis

    EasyOCR + regex pipeline achieving 30% accuracy gain.

    Publication
    Problem
    Handwritten tripsheets in logistics lead to high error rates in data extraction.
    Tech
    PythonEasyOCRRegex
    Impact
    Achieved 30% accuracy improvement through OCR and post-processing, published in IEEE Xplore.
    Read paperSource