David Smith

Data Scientist II

Data Scientist

• Developed the model and corresponding C++ ML library aimed at optimizing memory management with nanosecond constraints. • Contributed to open-source projects, increasing AI tooling and library compatibility with Windows ARM64 based architectures. • Developed ML training, validation and deployment infrastructure with Databricks and MLFlow. • Developed periodic and seasonal based data drift detection tools for increased model reliability based on the KL Divergence derived Population Stability Index. • Developed custom performance metrics to better represent ordinal label classifiers. • Reduced engineer workload by automating hyperparameter training and label derivation through the integration of FLAML and the development of a Spark friendly Jenk's Optimization Algorithm. Distributed Systems, Research and +7 skills

Software Engineer

• Costed, designed, proposed and led the implementation of solo and team projects to add new product features, increase maintainability or reliability to various Azure services. • Developed and contributed to open-sourced projects benefiting Azure and Azure customers. • Mentored and guided interns and new engineers to help them maximize team impact without sacrificing work-life balance. • Designed and implemented distributed system/micro-services (containerized apps orchestrated with Kubernetes) supporting 99.99% reliability, availability and P99 < 1s latency. • Fostered organization culture by planning and driving cross-team morale events for hundreds of employees. • Designed the integration between the Azure Bare-metal cloud service and VMWare, allowing specialized, high I/O servers, designed for Epic Systems workloads to be managed by VMWare and Azure. • Designed and developed a control plane for the GPU infrastructure used in training/running internal AI models. • Increased customer satisfaction and trust by providing live site problem investigation and mitigation. Docker, Distributed Systems and +3 skills

AI/ML Developer Intern

• Increased the generalization and data validation capabilities of a DVC based ETL pipeline. • Developed a species based named-entity recognition system with an extraction and tagging speed of 3.5 seconds per 1,000 abstracts. • Created and implemented a DynamoDB data model by initializing and updating tables via Lambda functions and defining the REST APIs through Velocity Template Language and Cloud Formation Research, Artificial Intelligence (AI) and +3 skills

AI/ML Developer Intern

• Explored and implemented topic modeling through various techniques such as Latent Dirichlet Allocation, Gensim Author Topic Modeling, and Unsupervised Learning. • Prepared and presented research proposals for a DARPA SBIR funded project to a company-wide audience of diverse technical backgrounds. • Reduced the project's main data set of 600+ features to 52 features while maintaining the existing neural network model's accuracy (within ~1%). This data set scaling allowed a decrease in training time from a few days to a few hours as opposed to the 5% reduction in model accuracy if the 600 features are simply trimmed to 52 random features. Natural Language Processing (NLP), Docker and +7 skills

Software Engineer Intern

• Implemented an ETL pipeline with DVC which included API integration, data-caching, web-scraping, data processing (for NLP) and data validation through Great Expectations. • Researched, trained and tuned advanced NLP models. • Mentored the more “green” interns in AI and software development techniques, and best practices. • Assisted CEOs, and subject-matter experts in the understanding and implementation of different software technologies. • Increased team productivity by refactoring agile processes. Research, Artificial Intelligence (AI) and +2 skills

AI Consultant

• Designed a latent diffusion tool based on the state-of-the-art Stable Diffusion work at the time that would operate in GPU and non-GPU environments and provide the same features as solutions produced by teams of researchers. • Implemented these PyTorch diffusion models end-to-end as a scalable docker-based system in a GCP environment through a Flask based API and a Celery based worker/node set up. • Developed the AI system so that different latent diffusion models/weights could be rolled in with minimal effort. • Characterized the energy efficiency of multi-node versus vertically scaled AI training methods. PyTorch, Docker and +2 skills