Experience Apr 2024 - Present
Machine Learning Engineer II
Guidewire
San Mateo, CA, USA • Full-Time
Leading enterprise ML systems across OCR-RAG-LLM optimization, claim summarization, geospatial hazard intelligence, digital twin simulation, and production AWS pipelines.
RLHFQLoRAQwenRAGAgentic AIAWSGeospatial CVDigital Twins
- • Automated prompt-prefix tuning for OCR-RAG-LLMs using RLHF, Bayesian optimization for LLM hyperparameter tuning, and CMA-ES for RAG hyperparameter tuning across five lines of business, reaching an average score of 0.875 across Levenshtein, F1-score, and ROUGE-L.
- • Finetuned and deployed a Qwen-3-4B claim-summarization QLoRA adapter with a 0.76 score, improving 6% across ROUGE-L, BERTScore, and domain-based evaluation after comparative analysis against GEMMA 3-4B QLoRA and base models.
- • Innovated an auto-annotation pipeline for community outline segmentation with hazard-based SatelliteMAE optical/SAR modeling, improving average Dice coefficient to 0.92 with late fusion for public safety buildings derived from news text and imagery.
- • Developed an agentic AI workflow with tool calling, chain-of-thought, tree-of-thought, and graph-of-thought reasoning across branched SFT DocLLM, OCR-RAG-LLM, and BDA systems to reduce false positive rates in Year Loss Tables for cyber risk modeling.
- • Architected digital twin simulation infrastructure to model complex cyber risks on network infrastructure, analyzing phase-shift signals and packet captures to simulate and predict multi-stage attack vectors.
- • Built batch and on-demand feature stores and pipelines with S3, Step Functions, DAGs, ECS services, and TeamCity CI/CD, using CodeLLaMA for HCL generation with evacc 81% and integrated Mann-Whitney U gate plus z-score temporal drift monitoring.