At Accenture, I deployed a Gemini 2.5 Flash-Lite image-classification pipeline for a logistics client, improving parcel-type identification accuracy from 90% to 97% and reducing manual review workload by about 60%.
I architected Vertex AI data and ML pipelines on GCP that process about 100K images per week. I also curated a 15K+ ground-truth image dataset and used drift detection to help maintain accuracy across changing hardware and operating conditions.
For my Agentic Financial Research Assistant project, I designed a multi-agent RAG platform with an 12-node LangGraph state machine. Its fast-path optimization reduced LLM API calls by 70%, and its OpenSearch hybrid retrieval pipeline combined BM25 and HNSW search with reranking.
I also built a spatio-temporal taxi demand forecasting pipeline using H3 spatial aggregation and LightGBM quantile regression. It achieved 22.49% WAPE on a five-week out-of-time holdout, with uncertainty estimates for capacity planning.
