Senior Data Scientist · Applied ML Engineer
Ali Abul-Hawa, PhD
Selected work across multilingual retrieval, large-scale Arabic NLP, visual anomaly detection, local document Q&A, reliable software systems, and earlier control/energy research.
Selected work
Systems, analyses, and technical case studies
Arabic Social Media NLP at Scale
A 700M+ record social-data pipeline spanning streaming ingestion, MongoDB deduplication, automated/spam-source filtering, dialect-aware Arabic normalization, embeddings and supervised classification.
Open case study → Professional work · NLP · Information retrievalProduction Semantic Retrieval
MPNet-family sentence representations and Elasticsearch integrated into an established multilingual product, with model selection grounded in a nine-variant embedding comparison rather than an isolated embedding demo.
Open case study → LLM systems · Retrieval · EvaluationLocal Document Q&A
Hybrid lexical/dense retrieval, source-aware generation and tracing, including a controlled experiment where adding a cross-encoder reduced strict Hit@5 from 0.90 to 0.60 and was therefore left disabled.
Open case study → Team project · Machine learning · Computer visionIndustrial Visual Anomaly Detection
PatchCore, autoencoders and DINOv2/DINOv3 nearest-neighbour baselines on MVTec AD, with real localization outputs and a fair 15-category comparison where DINOv2 improved image-level F1 over PatchCore.
Open case study →Independent engineering
Software projects
HomeDrive
Android-to-Windows backup built around authenticated pairing, explicit verification, capacity preflight, atomic writes, reconciliation, and fail-closed deletion safety.
Open case study → Full-stack · Learning systemGerman Master
Lexeme-centric, multi-part-of-speech practice with deterministic task generation, Postgres, content enrichment, offline PWA behavior, queued synchronization, and shared user history.
Open case study → Algorithms · AndroidSudoku Generator Engineering
Irregular Sudoku generation using backtracking, MRV, constraint handling, uniqueness checks, performance tuning, and pipeline reliability.
Read engineering article →Earlier work & research
Computer vision, control, and energy systems
100-class Food Recognition
UECFOOD100 classification with ImageNet-pretrained VGG16, Keras/Theano, bottleneck features, fine-tuning, augmentation, GPU training and a phone-to-model calorie-per-serving lookup.
Open historical case study → PhD research · Modeling · Control · EnergyResearch & Engineering
Dynamic PEM fuel-cell/load modeling and nonlinear control under variable demand, followed by photovoltaic sizing and life-cycle economic analysis.
Open research page →