At CDM Smith, I develop document-matching pipelines that cut unmatched research documents from 858K to 577K and added 787K filings for batch and streaming use.
I built RAG over research reports and filings, combining vector and keyword search with reranking and source citations. Evaluation on 500 labeled questions showed a 19% lift in answer accuracy over keyword-only search.
I fine-tuned BERT and T5 models for classification, summarization, and entity extraction, reaching 0.91 F1 and cutting manual tagging by 40%. I also developed an LLM workflow that matched analyst reviews 92% of the time.
Previously, at Enlightened Infosystems, I automated a database-to-prediction pipeline that reduced monthly refresh time from 3 hours to 20 minutes. On an open-source project, I built hybrid search for SEC filings that lifted recall@10 from 0.71 to 0.86.
