Youssef Mostafa

Fine-tuned and quantized SLMs/LLMs (Llama, Qwen, DeepSeek) for in-car NLU, speech recognition, and agentic assistants under strict latency (<200ms) and memory (<2GB) budgets, managed full training-to-deployment lifecycle on Android test vehicles, achieving 92% intent accuracy on-device Architected GraphRAG and hybrid RAG pipelines for the BMW Intelligent Personal Assistant; designed retrieval with semantic search (dense embeddings + BM25 sparse retrieval), reranking, and query-aware chunking to reduce retrieval latency by 60%; implemented confidence-calibrated response generation, reducing hallucination rate by 40% through structured evaluation and guardrails Optimized distributed training via Ray/KubeRay on AWS (EKS, S3, ECR) achieving 5–10x convergence speedup; built centralized GPU clusters with Karpenter autoscaling, enabling 550+ developers across 60+ ML use cases; self-hosted large-scale LLMs (DeepSeek, Qwen-72B) on managed GPU infrastructure; contributed to platform architecture showcased at Ray Summit 2025 Developed battery state prediction models from vehicle sensor telemetry with physics-informed feature engineering and temporal modeling, handling lab-to-road distribution shift; built occupant monitoring system using multi-task vision with knowledge distillation for edge deployment Built CAIP Python SDK abstracting ML workflows behind clean interfaces with Docker packaging and backward-compatible versioning; implemented inference monitoring, prediction drift detection, and LLM evaluation pipelines (LangSmith, custom trace-based debugging) for safety-critical deployment

Fine-tuned document layout analysis models for legal document segmentation (tables, headers, nested clauses); trained domain-adapted NER on legal corpora with custom entity schemas, cross-page coreference resolution, and nested entity handling, achieving 89% F1 on complex multi-page contracts Built multi-stage extraction pipeline with confidence calibration and uncertainty propagation; designed multi-agent architecture (LangChain, LangGraph, AutoGen, CrewAI) with learned routing for document-type-specific strategies, reducing manual review time by 65% Deployed ML microservices on Google Cloud Run with cold-start optimization and memory-efficient model serving (vLLM, ONNX Runtime); built structured output schemas and inter-service data contracts for production reliability

Developed multivariate time series forecasting models for manufacturing process optimization (BMW, Airbus, Fraunhofer) with physics-informed feature engineering and causal feature selection; reduced defect rates by 28% through predictive intervention Built unsupervised anomaly detection systems using reconstruction-based and density estimation approaches with adaptive threshold selection under label scarcity; validated via synthetic fault injection and expert-in-the-loop evaluation Trained digital twin models with continual calibration against production data and distribution drift monitoring; automated ML pipelines with MLflow, Docker, and GitHub Actions for reproducible training and deployment

Built REST APIs across multiple stacks (Python/FastAPI, Laravel, NestJS) with event-driven webhook architecture for payment integration (Paddle, PayPal); developed React frontends and internal tooling; owned code review and repository management across concurrent projects

Developed Fintech, E-commerce and SAAS Applications

Developed privacy-preserving multi-face detection and recognition system for thermal/IR imaging, handling domain gap between visible-spectrum and infrared modalities; achieved 95% accuracy through custom domain adaptation and data augmentation on IR-specific noise characteristics Built NLP pipeline for summarization, contextual reference resolution, text-to-speech integration, and background noise removal; combined techniques from 15+ research papers into production-grade systems

Worked on creating Website for client companies transitioning client requests to user stories and then realizing them optimizing backend and database functionalities testing stubs and doing unit test

Recreated the universities main website designed a more interactive UI/UX handled and tested endpoints followed agile and scrum workflows in a team of 6

have successfully graduated from the React Advanced Front End Web Development Nanodegree class

have successfully graduated from the Advanced Web Development Nanodegree with top score