Ai Developer

Built and fine-tuned a multilingual BERT-based sentiment-analysis pipeline for Chinese and global media, then modeled its impact on key Chinese economic indicators. Collaborated with economists to turn signals to KPIs, adding safeguards for source drift & anomalous spikes.

Performed domain-adaptive pretraining of LLaMA on language prompts derived from structured NACC dementia data, followed by supervised fine-tuning for differential diagnosis of dementia. Fine-tuned using 4-bit quantization approach (QLoRA) reducing GPU memory usage by 50%. Developed a RAG pipeline to generate summaries grounded in PubMed journals, providing citation-backed explanations for LLaMA’s dementia diagnosis predictions, improving interpretability and clinical trust. Engineered an LLM-as-a-judge evaluation framework to build metrics to assess groundedness, contextual relevance, and hallucination in RAG-generated summaries. Then used it as a quality gate before pushing updates. Deployed a Dockerized Streamlit UI & inference stack on AWS with monitoring & structured logs for error rates.

Delivered AI-driven analytics engine utilizing language models & heuristics to forecast blockchain project outcomes.

Constructed MLOps pipelines to track IoT-based refrigeration metrics across 5 500+ Walmart stores, reducing food spoilage and cutting data-retrieval time by 70 % through optimized SQL queries. Led a cross-functional data quality initiative for store sensor data, improving valid data coverage from 37% to 80%. Deployed the enhanced data pipeline and ML models to production on Google Cloud (BigQuery), resulting in more accurate insights, and received a Walmart 'Bravo' Award for this effort. Built predictive maintenance models to forecast forklift breakdowns, delivering $1M+ in cost savings. Implemented model interpretability techniques (LIME, SHAP) increasing trust & adoption for asset health models.