Jahnvi Shah

Data Science Intern @Locobuzz

Built, fine-tuned, and deployed a 3B LLaMA-3 for customer-response generation across 40+ channels; stood up a 98% QA gate for tone, factuality, and placeholder safety.Engineered a modular LLM pipeline (data ingestion → prompt templates → model configs) and turned results into Matplotlib/Streamlit dashboards for weekly stakeholder readouts (clear ship/hold decisions).Cut hallucinations by 35% via prompt design (few/zero-shot) and evaluation on a 500-prompt suite; tracked F1/accuracy with versioned datasets.Optimized inference with 4-/8-bit quantization and KV-cache reuse to balance accuracy, latency, and GPU memory.