Mark Kien

AI / ML Engineer specializing in Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), model fine-tuning, pretraining workflows, and reinforcement learning. Experienced in designing generative AI applications, intelligent assistants, AI agents, and machine learning solutions. Strong expertise in Python, deep learning frameworks, transformer architectures, vector databases, and production AI system development.

AI / LLM Engineer at Ineffable Intelligence (2025 – 2025)

• Designed LLM-powered applications combining retrieval, embeddings, structured generation, and backend services.
• Architected RAG workflows for enterprise knowledge discovery, including ingestion, chunking, retrieval, reranking, context construction, and grounded generation.
• Reduced fine-tuning GPU memory consumption by approximately 60% using QLoRA, allowing larger models to be trained on the available GPU infrastructure.
• Developed evaluation workflows for answer quality, retrieval relevance, latency, reliability, and model cost.
• Strong production experience with MCP and Hermes Agent, building and orchestrating tool-using, multi-agent workflows with API/DB integrations, RAG, memory, planning, error recovery, and production monitoring.

LLM / Data Engineer at Imobisoft (2022 – 2024)

• Designed a distributed Transformer pretraining pipeline using PyTorch, FSDP, mixed precision, and multi-GPU training, processing more than 1.2B training tokens per day.
• Fine-tuned open-source LLMs using SFT, LoRA, QLoRA, and PEFT for domain-specific instruction-following tasks.
• Reduced unnecessary LLM calls by approximately 28%, lowering inference costs for high-volume agent workloads.
• Experimented with reward-model-based training and reinforcement-learning workflows for specialized generation tasks.

AI Engineer at Dynamo AI (2021 – 2022)

• Developed Transformer-based NLP systems for document classification, information extraction, semantic similarity, and question answering.
• Fine-tuned BERT and other Transformer models on domain-specific datasets.
• Built automated benchmark pipelines for evaluating model quality across instruction following, reasoning, extraction, summarization, and domain-specific tasks
• Improved document classification F1 score from 82% to 93% through data augmentation, fine-tuning, and hyperparameter optimization.

M.S. in Computer Science – Stanford University (2018 – 2020)

Bachelor of Science in Computer Science – California State University (2014 – 2018)