Projects

Built an agentic AI assistant layer into a React-based fiber-network planning application, letting planners query and reason over map and network data in natural language.

• Designed a hierarchical tool-selection system that routes user intent across a catalog of 48 tools, with session-level tool caching for responsive multi-turn use.
• Implemented MCP servers on Podman-based local infrastructure with streamed chat responses and structured tool results.
• Added dual-provider inference within one architecture: local Ollama models and cloud Claude, selectable per deployment.
• Kept the assistant deliberately read-only and integrated with existing map and network-data workflows; built with Claude Code.

Built field-ready computer vision workflows for construction-site and infrastructure imagery.

• Designed segmentation pipelines for trenches, ducts, fibers, and fittings in construction-site imagery.
• Handled lighting variation, occlusion, dirt, shadows, and ambiguous material boundaries during evaluation.
• Converted models to CoreML and built iPhone prototypes for on-device visualization and testing.
• Connected annotation strategy, model evaluation, failure analysis, and deployment preparation in one workflow.

Developed privacy-conscious local AI workflows for documents, images, and knowledge extraction.

• Built local LLM-based workflows for document understanding, semantic search, and knowledge extraction without external cloud inference.
• Extracted structured information from complex technical documents, PDFs, and tables.
• Combined retrieval, vision-language models, and RAG-style approaches for automated document analysis.
• Built CLIP-based semantic image search using natural language queries, embeddings, and similarity-based clustering.

Developed machine learning approaches for automotive body production, 3D deformation learning, and fast surrogate prediction from FEM simulation data.

• Created large-scale FEM training datasets for sheet metal forming where no suitable dataset existed.
• Worked with mesh autoencoders, implicit neural representations, signed distance functions, and geometric deep learning.
• Built surrogate models for geometric deviation, thickness distribution, and thinning on high-resolution shell meshes.