Gerard Christian Kuotu

Led end-to-end development and launch of a Voice AI English Tutor within an LMS; established Model Context Protocol layer to enable low-latency bi-directional voice and structured data exchange. Built an event-driven MLOps architecture using AWS SQS and Fargate; developed a Hybrid RAG system with vector search and metadata filtering; trained and deployed multilingual TTS models. Designed an automated LLM-as-a-Judge evaluation framework automating 90% of model QA and reducing rating variance by 20%. Led cross-functional onboarding for Product and Data Science teams and defined core LLM architecture and MLOps standards.

Led design and deployment of a HIPAA-compliant, multi-agent Voice AI for clinical applications; architected and scaled a Hybrid Retrieval-Augmented Generation system using PgVector and HNSW-based indexing with a TinyBERT reranker to reduce hallucinations and boost retrieval precision. Integrated Generative AI Guardrails and custom safety checkers; optimized LLMs with vLLM, FlashAttention, and dynamic batching on AWS EKS, increasing GPU utilization by 30% and cutting average inference latency by 50%. Built distributed fine-tuning pipelines for open-source LLMs (Llama, Gemma) using QLoRA and Instruction Tuning on SageMaker; implemented RLHF with PPO. Implemented MCP server integration with LiveKit and Tavily Search; established Prompt Engineering, HITL feedback loops, and an automated evaluation pipeline with LLM-as-a-Judge.

Architected Copilot Factory Framework (CFF), an internal LLM orchestration system to rapidly prototype and deploy domain-specific chatbots for financial analysts. Built end-to-end MLOps pipeline using Kubeflow, Terraform, and CloudFormation on AWS EKS; containerized a BERT-based NER model; implemented BERT-based topic modeling (Gensim) for 10M+ records annually; established evaluation and monitoring with Weights & Biases; improved image recognition accuracy to 98% via optimized computer vision pipeline (OpenCV/TensorFlow).