Richard Chiang

A custom search algorithm leverages CLIP embeddings to identify home attributes like roof material from listing images. Reciprocal rank fusion scores image-text relevance to retrieve results.

A CNN is fine-tuned in PyTorch Lightning to classify roof age from satellite imagery. This model automates risk assessment for underwriters.

Additionally, an end-to-end satellite imagery data pipeline was developed to train these models. The pipeline scrapes thousands of roof images from… Show more

A custom search algorithm leverages CLIP embeddings to identify home attributes like roof material from listing images. Reciprocal rank fusion scores image-text relevance to retrieve results.

A CNN is fine-tuned in PyTorch Lightning to classify roof age from satellite imagery. This model automates risk assessment for underwriters.

Additionally, an end-to-end satellite imagery data pipeline was developed to train these models. The pipeline scrapes thousands of roof images from mapping APIs, cleans and annotates the data. Show less

This system enhances the digital journaling process through conversational interactions. Using retrieval augmented generation, it facilitates brainstorming, note-taking and collaborative editing of journal entries.

At the core is a hybrid search architecture combining BM25 and dense retrieval for fast retrieval from thousands of journal entries. Temporal understanding allows filtering by extracted relative dates and time-weighted scoring surfaces recent topics.

The model was… Show more

This system enhances the digital journaling process through conversational interactions. Using retrieval augmented generation, it facilitates brainstorming, note-taking and collaborative editing of journal entries.

At the core is a hybrid search architecture combining BM25 and dense retrieval for fast retrieval from thousands of journal entries. Temporal understanding allows filtering by extracted relative dates and time-weighted scoring surfaces recent topics.

The model was trained on augmented journal text prompts and optimized for low-resource scenarios through quantization. Natural language interactions integrate the system seamlessly into any journaling workflow. Show less