Ruchi Bhanot

COURSES COMPLETED:
Fundamentals of Generative Artificial Intelligence (AI) 
Applied Machine Learning
Statistics for Data Science
Research Design and Applications for Data and Analysis
Introduction to Data Science Programming

SELECTED PROJECTS

DATASCI267: Fundamentals of Generative AI
Abstract: Designed, compared and evaluated open-source and vendor LLMs to build a production-oriented, multi-model RAG framework using Mistral, OpenAI and Cohere models with model-specific prompt engineering and separate training and RAG pipelines. Benchmarked models using lexical, semantic and LLM-as-Judge evaluation methods, and finally, optimized model-specific prompts and RAG components and analyzed production deployment/LLMOps tradeoffs, including cost, latency, scalability, privacy, compliance and availability. 
Models used: Mistral's llm_mistral_model, Cohere's CohereChatTemplate, OpenAI's gpt-5.4-mini with ReRanker, ReWriter used for improving query focused hyper parameters in the pipeline.
Evaluations tools: Lexical (BLEU, ROUGE, BERT), Semantic (Cosine Similarity), LLM-as-Judge (RAGAS: context_precision, context_recall, precision_score, faithfulness, answer_relevancy).

DATASCI207 - Applied Machine Learning
Title: Predicting Wildfires in California Using Machine Learning
My contribution included supporting the EDA, data shuffling and splitting, performing the logistic regression analysis. Doing the literature review, related work research, co-writing the methodology (Logistic Regression), Results and Discussion sections. 
Abstract:
Wildfire ignition prediction is a critical challenge for fire protection agencies seeking to allocate resources before fires start. This project applied supervised binary classification to predict whether a wildfire will ignite on a given day in California using over 14K daily weather observations spanning 1984 to 2025. 
Model input:  23 weather and temporal features including temperature, precipitation, wind speed, drought duration, and seasonality indicators. 
Models tested: Logistic regression, XGBoost, Random Forest, and a feedforward neural network (FFNN), against a majority class baseline. 
Results: All models substantially outperformed the baseline (66.82% test accuracy, zero fire detection), with XGBoost as our top performing model. Seasonal subgroup analysis revealed that all models perform well in summer driven by high base fire rates but fail to detect winter and spring fires where weather features alone provide insufficient discriminating signal. These findings suggest the predictive ceiling of tabular
weather data has been reached, and that future improvements require spatially resolved data, more observations and additional environmental features.
Git repo maintained: eda_and_preprocessing_Ruchi.ipynb (ruchit branch) <https://github.com/alvarolaham-ucb/207_final_project_Alvaro_Jake_Ruchi_…;