Nihad Hassan

Maintained a central ML gateway (FastAPI) routing inference traffic across multiple NLP and multimodal services using RabbitMQ and Kubernetes, supporting millions of requests per day across high-throughput, multi-language workloads. Led end-to-end deployment of NLP models (classifiers, paraphrasers, grammar checkers, translators) across multiple GCP clusters and environments, enabling safe rollouts and rollbacks using ArgoCD and Datadog. Built an automated CI/CD pipeline using GitLab CI to promote fine-tuned Vertex AI models from model registry to production, integrating LiteLLM-based cost observability to track and optimize LLM inference spend, increasing developer efficiency by 70%. Designed resilient LLM inference pipelines with Langfuse-driven prompt versioning and automated cross-vendor fallback, ensuring graceful failover during rate-limit and availability incidents and reducing user-facing failures during traffic spikes. Led complex embedder-discriminator deployments to support personalization research, collaborating with data, backend and research teams to safely productionize experimental models and accelerate research-to-production timelines. Implemented batch processing for AI content detection workloads, improving throughput and reducing required Kubernetes pods, resulting in meaningful infrastructure cost savings.

Trained and deployed a cascade model solution combining Semantic lexicon and LSTM for Brickcode prediction on 1.3 million rows of data as a serverless Azure Function App (HTTP trigger), leveraging Azure ML Studio. Optimized inference time by 75%, significantly enhancing prediction accuracy and deployment efficiency. Developed a nutrient attributes extraction solution using Azure Cognitive OCR, LangChain, and AzureOpenAI to extract, normalize, and rephrase nutrition information from images, employing GPT-4 prompt engineering and regex for structured data processing, and deployed it as a serverless Azure Function App. Designed and deployed an image validation and enhancement pipeline as an Azure Function App, processing user-provided image URLs through validation checks, enhancement, and lossless compression, then serving the optimized images via Azure blob storage URLs. Developed a PDF to DOCX converter utilizing PaddleOCR, optimized to run on standard servers instead of GPU-based servers with minimal inference time difference, reducing costs by 80% and enhancing document management and accessibility.