Adaptive ML pipelines
Designed online feature pipelines with 200+ engineered signals and past-only transformations to maintain consistency between historical and live inference.
Omar Elkobrossy
Software Engineer | Machine Learning & Systems Focus
Building real-time systems, ML pipelines, and data-driven applications.
Computer Science student with hands-on experience building production software, adaptive ML systems, and end-to-end data pipelines.
index
I'm a Computer Science student and software engineer focused on machine learning systems and reliable software infrastructure. I enjoy building systems that handle messy real-world data, from adaptive ML pipelines to production software and computer vision applications.
work
-
2023 — present Lexington, KY
Software Engineer
Bluegrass Integrated Communications
- Reduced customer and production data processing time by 46% by building an event-driven pipeline that ingests and consolidates data from multiple internal sources.
- Automated document and layout generation by developing a placeholder-based templating system used across internal workflows.
- Engineered middleware to validate, enrich, and cross-reference customer submissions with internal production databases.
- Maintained operational continuity by building a GUI tool with manual override capabilities and error handling for failure scenarios.
-
2021 — 2022 Alexandria, Egypt
Software Developer
Vortex Academy
- Built and programmed ROV mobility systems using C++ and Arduino, contributing to a first-place competition finish.
- Developed Python-based algorithms for technical and competitive programming challenges.
- Designed, tested, and iterated on ROV subsystems to improve reliability during competition scenarios.
- Implemented computer vision functionality to support object detection and navigation tasks.
systems & engineering
Regime-aware model adaptation
Built drift-triggered retraining using PSI and historical state analysis so models adapt automatically to changing distributions.
46% faster production workflows
Reduced processing time by 46% through event-driven architecture and consolidation of fragmented internal systems.
Model observability & diagnostics
Engineered monitoring systems for calibration drift, prediction collapse, gradient health, GPU usage, and failure recovery.
projects
// 001
github ↗
Chronoflow
Adaptive ML system for non-stationary financial time series.
- Leakage-safe feature pipeline with 200+ features, past-only transformations, and online statistical scaling.
- Regime-aware parameter tuning (FAPT) using Wasserstein similarity between historical states.
- Event-driven retraining via PSI drift detection and walk-forward validation.
- Joint model + strategy optimization with Optuna, distributed on AWS Batch.
// 002
CheXpert Classifier
Multi-label chest X-ray classification with ConvNeXt.
- Patient-level stratified splits preserving rare pathology distributions across train/val.
- Weighted loss + soft-target handling for class imbalance and uncertain medical labels.
- Training diagnostics for gradient health, calibration, prediction collapse, and GPU utilization.
- Resumable training with separated model checkpoints and training-state recovery.
stack
languages
- Python
- SQL
- C++
ml
- PyTorch
- XGBoost
- Scikit-learn
- Optuna
- Tensorflow
- Keras
infrastructure
- Docker
- AWS
- Linux
- REST APIs
focus areas
- ML Systems
- Real-Time Data
- Computer Vision
contact
Feel free to reach out regarding software engineering, ML systems, or collaborative work.
- email omarelkobrossy@yahoo.com
- phone +1 859 420 0509
- linkedin linkedin.com/in/omar-elkobrossy ↗
- github github.com/omarelkobrossy ↗