Sherry Shi

Scalable Data Ingestion Pipeline - Built and deployed a 24/7 ingestion service on a GCP Linux VM, processing and persisting 1.4M+ normalized entities into MongoDB for downstream product and operations use. - Designed a fault-tolerant pipeline with queue scheduling, checkpointing, and idempotent writes for safe restarts in long-running workloads. - Optimized sustained throughput under external API constraints using async concurrency and rate limiting, improving processing speed by ~4.0×; added structured, rotating logs for observability. - Evaluated the Hugging Face transformer model for profile attribute inference as an ML-based enrichment step. LLM Retrieval-Augmented Q&A System - Built a Python web application with chat UI enabling natural language queries over Excel marketing reports using a RAG pipeline. - Reduced embedding indexing time ~37.0× by implementing batched embedding API calls. - Designed a configurable document processing pipeline including chunking, vector embeddings, and semantic retrieval. - Documented system architecture, APIs, and implementation details across projects; wrote unit tests for each module to ensure reliability and maintainability.

- Built a GenAI evaluation tool with Gradio UI, measuring response quality, latency, and token usage across multiple model providers. - Designed an extensible object-oriented adapter architecture to standardize provider inputs and outputs. - Implemented parallel request execution with concurrency control, achieving ~3.0x throughput improvement under API rate limits. - Added privacy-safe JSONL logging with prompt fingerprinting for offline analysis.

Dean’s Honors for 5 semesters