ML Engineer | Deep Learning · Computer Vision · Scalable Data Pipelines
I specialize in bridging the gap between prototypes and production. Whether it's training AI models with high accuracy, engineering robust data pipelines, or automating hours of manual workflows into minutes, my focus is on building scalable, real-world systems.
- Applied AI & ML: Training models and deploying them as production-ready REST APIs with a focus on explainability and out-of-distribution (OOD) robustness.
- Backend & Data Engineering: Building fault-tolerant architectures and scalable ETL pipelines that handle heavy data efficiently.
- RPA & Automation: Developing resilient, multi-threaded automation scripts that drastically reduce operational workloads.
- AI & Data:
TensorFlowScikit-LearnOpenCVPandasDask - Backend:
PythonFlaskSQLPostgreSQLREST APIs - Automation:
SeleniumAppiumMultithreadingUiAutomator2 - Tools:
DockerGit
| 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. |
| 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. |
| listing-pilot | Config-driven mobile automation suite managing ~1,000 active listings across C2C marketplaces. Built with Appium and Python, featuring overlap detection, multi-ID fallbacks, and real-time Telegram alerting. Reduced daily manual workload by over 90%. |
| popcorn-wagon | Hybrid movie recommender engine built with Dask/Pandas for scalable ETL and Spotify Annoy for sub-millisecond similarity search. |
| portal-cleaner-ultimate | Modular RPA suite featuring a custom local test harness for offline ERP simulation. Engineered with fault-tolerant retry logic, slashing manual operational workloads by over 90% (4–6 hours to under 30 minutes). |