Applied ML Engineer | Computer Vision · LLM Systems · Production ML
I specialize in bridging the gap between AI prototypes and production. Whether it's training robust Computer Vision models, designing Retrieval-Augmented Generation (RAG) pipelines, or building high-throughput streaming architectures, my focus is on engineering scalable, real-world ML systems.
- Applied AI & ML: Training deep learning (CV) and generative AI (LLM/RAG) models, deploying them as production-ready microservices with a focus on out-of-distribution (OOD) generalization and explainability.
- MLOps & Data Pipelines: Building automated training pipelines, CI/CD for model deployment (GitHub Actions, Docker), and out-of-core streaming architectures (Dask, drift detection).
- Backend & Systems Engineering: Architecting fault-tolerant backend workflows, high-concurrency data ingestion pipelines, and resilient multi-threaded systems.
- ML & Deep Learning:
PyTorchTensorFlowScikit-LearnTransformersSentence Transformers - Computer Vision & NLP:
OpenCVImage ProcessingRAGChromaDBSpotify Annoy - Data & Backend:
PythonFastAPIFlaskPostgreSQLDask (ETL)PandasSQL - DevOps & Infrastructure:
DockerGitHub Actions (CI/CD)GitLinux
| Project | Description |
|---|---|
| wbc-analyzer | End-to-end WBC classification system deployed as a Flask REST API. Features a custom architecture (DenseNet121 + WBCAttention + MedSwish). Incorporates inference-time domain adaptation achieving an 89.05% out-of-distribution (OOD) accuracy (+32.09 pp boost). Includes a multi-modal LLM agent (GPT-4o & Gemini) for clinical XAI insights. Preprint published on ResearchGate. |
| rag-project-assistant | Hybrid-source RAG system answering questions about my portfolio projects. Combines AST-extracted code structure with curated documentation, using similarity-threshold gating (ChromaDB, L2 < 1.40) to prevent hallucination on out-of-scope queries. Deployed on Hugging Face Spaces (Docker, FastAPI, Groq Llama 3.3 70B) with rate-limited public API and a live chat widget. |
| kinematic-action-recognition | Full end-to-end ML pipeline on 10 GB motion-capture sensor data. Features out-of-core ingestion with Dask, real-time streaming with drift detection (81 windows/sec), and a LightGBM/RandomForest ensemble achieving 0.94169 accuracy on Kaggle. |
| popcorn-wagon | Hybrid movie recommender engine built with Dask/Pandas for scalable ETL and Spotify Annoy for sub-millisecond similarity search. Integrates collaborative filtering (SVD) and content-based filtering. |