Evan Carlson Email and Phone Number

- Improved existing text classification model using QLoRA, improving precision by 43.33% and recall by 13%. - Technologies: Hugging Face, Python, PyTorch.- Developed domain-adapted LLMs to enhance downstream tasks, achieving up to 10% performance boost. - Technologies: Hugging Face, Python, PyTorch.- Implemented techniques such as progressive layer freezing, data filtering/cleaning, and context size experimentation for pre-training LLMs. - Technologies: Hugging Face, Python, PyTorch.- Engineered RAG application on both structured and unstructured data, allowing customers to ask complex questions about their historical policy information then receive an answer with supporting documentation. - Technologies: Hugging Face, LangChain, Pinecone, Python, PyTorch.- Trained generative AI models for automated information extraction from transcripts, streamlining claims handling. - Technologies: Hugging Face, Python, PyTorch.- Optimized LLM inference for information extraction using quantization, distillation, JIT tracing, and TensorRT compilation with ONNX. Reduced RoBERTa size and latency by 50% and T5 size by 70% and latency by 50% with no performance degradation. - Technologies: Hugging Face, ONNX, Python, PyTorch.- Debugged, fixed, and managed Docker images, ensuring seamless integration and deployment across teams.- Reduced public Llama-2 13B model’s CPU latency by 67% using GGUF, reducing deployment costs. - Technologies: C++, GGUF, Python.-Profiled and optimized large Transformer models, decreasing training time via model parallelization, quantization, and memory format conversion for PyTorch. Increased average GPU utilization. - Technologies: NVIDIA NSight, Python, PyTorch.- Templated model deployment process, fostering uniformity and efficiency across the organization. - Technologies: Jenkins, Python, PyTorch, Tensorflow

- Spearheaded NLP R&D pipeline, automating model training and validation and improving model deployment efficiency. - Technologies: Python, PyTorch.- Reduced DistilBERT inference time by approximately 50%, maintaining high-quality results with less than 1% loss in accuracy using dynamic quantization. - Technologies: Python, PyTorch.- Orchestrated multi-phase upgrade of internal Postgres databases from 9.6 to 13.3, resulting in three-fold faster query times. - Technologies: Postgres.- Integrated billing system, providing finance team with greater data control and flexibility. - Technologies: Golang.- Oversaw data mining services, ensuring strict data quality and swift collection from various sources. - Technologies: Ruby.- Profiled database queries and checked for bottlenecks in APIs to increase application speed and efficiency. - Technologies: Golang, Postgres, Ruby.- Contributed to the creation of general API features, ensuring product security and functionality. - Technologies: Golang, Python, Ruby.

- Led database architecture rewrite, enabling real-time event-driven processes at scale. - Technologies: GCP, Kotlin, NoSQL, Python, Swift.- Co-led ecosystem simulation program, facilitating analysis of smart device behaviors. - Technologies: Python, SimPy.- Configured CI/CD pipeline for mobile development. - Technologies: Jenkins.

- Collaborated in team to design and develop customer facing web application. - Technologies: CSS, HTML, Node.js, Python.- Designed and developed customer facing iPhone application. - Technologies: Swift.- Designed and developed in-house and customer facing AI applications including object detection, gun detection, and face detection. - Technologies: Python, PyTorch, Swift.- Provided customers with an open forum for questions regarding strategicplans, best practices, and advice on upgrades, special concerns, and ad hoc consulting for software implementation, integration, training, and change requests.

- Implemented highly efficient and automated reports for internal and external clients. - Technologies: C#, Excel, Python, Report Builder, T-SQL, VB .NET, Visual Studio.- Regularly and proactively communicated status, risks, and other issuesthrough JIRA.- Analyzed, debugged, and improved previously written programs to increaseeffectiveness, efficiency, and security.