Research & experience.

Summer 2026 · LLM evaluation and skill-based agent systems.

• Designed controlled experiments and refined scoring rubrics for an LLM autorater, achieving over 90% repeated-evaluation score consistency and over 90% alignment with human ratings.
• Designed a graph representation of agent skills to encode sequencing and dependencies for relationship-aware retrieval, orchestration, and skill-library health checks.

Computer vision and 3D method evaluation, model adaptation, and training iteration.

• Built controlled image-matching and 3D reconstruction benchmarks under low-texture, repeated-pattern, sparse-view, and illumination-shift conditions to guide production method selection.
• Integrated FiLM-conditioned camera-orientation cues into a Swin Transformer segmentation model and trained inspection models through an internal active-learning pipeline.

LLM applications, human-AI workflows, and industry prototypes.

• Built a customer-support LLM backend prototype with retrieval, tool orchestration, transaction workflows, and human handoff.
• Prototyped inspection applications with iOS and web interfaces and deployed machine-learning and custom algorithm modules on AWS using SAM and Lambda for industry pilots.

• Led six industry research projects developing and deploying deep-learning inspection systems for concrete and steel structures, integrating visual recognition into field inspection and near-real-time reporting.

Geometry-Grounded Synthetic Data for Visual Recognition

Parameterized 3D defect models generate rendered imagery, pixel-level masks, and physical defect attributes without manual annotation.

• Developed a synthetic-data framework that derives supervision directly from controllable defect geometry.
• Generated 72K synthetic crack images and evaluated transfer across three public real-image segmentation benchmarks; realism enhancement improved zero-shot IoU by 3.7–13.6 points.
• Developing a CAD-referenced extension for multi-view appearance and geometry representations, matching, and few-shot asset recognition.

Vision-Language Learning for Visual Reporting and Reasoning

Vision-language models and datasets for generating and evaluating reports from visual observations.

• Co-developed VisualSiteDiary, a detector-free vision-language Transformer for construction photolog captioning and image retrieval (equal contribution).
• Developed VL-Con, a construction vision-language dataset and few-shot prompting framework using in-domain visual-text demonstrations.
• Evaluated single- and multi-agent vision-language workflows for inspection reporting, including model-generated self-assessment to identify low-quality outputs.

LLM Evaluation and Agent Workflows

Controlled experiments and rubric iteration for an autorater of troubleshooting hypotheses; agent skill-graph design.

• Compared scoring formulations and rubric-grounded prompts, using disagreement analysis to refine the evaluator.
• Achieved over 90% repeated-evaluation score consistency and over 90% alignment with human ratings on the evaluated tasks.
• Designed a graph representation of skill sequencing and dependencies for relationship-aware retrieval, orchestration, and library checks.

Vision Model Adaptation and Benchmarking

Controlled benchmarks and model adaptation for infrastructure and facility imagery.

• Compared image-matching and 3D reconstruction methods under low-texture, repeated-pattern, sparse-view, and illumination-shift conditions.
• Integrated FiLM-conditioned camera-orientation cues into Swin Transformer segmentation and extended inspection-specific object classes.
• Trained and iterated on domain-specific recognition models using the company's existing active-learning pipeline.

Applied AI Prototypes and Human Handoff

Customer-support prototype integrating retrieval, tools, backend transactions, and human escalation.

• Built the LLM backend and agent-console handoff for an omnichannel support prototype.
• Developed conversation-analysis workflows and an embedding-based classifier for identifying effective service utterances.
• In separate inspection projects, built iOS and web interfaces and deployed modules on AWS using SAM and Lambda for industry pilots.

GPA: 4.0

• Research: multimodal AI, computer vision, 3D vision, and geometry-grounded synthetic data.
• Teaching Assistant, Computer Vision (Spring 2024); Government Fellowship for Studying Abroad (2024–2026).

GPA: 4.0