Tripti Garg

Designed and implemented retrieval-augmented generation (RAG) systems for structured and rule-dense domains including defense regulations and analytical document workflows. Developed structure-aware document chunking strategies and embedding pipelines using Sentence Transformers and FAISS to improve retrieval relevance (optimized Recall@K). Performed supervised fine-tuning and domain adaptation of transformer-based models using Hugging Face and LoRA/PEFT to improve response consistency on domain-specific tasks. Built grounded evaluation frameworks incorporating citation enforcement and retrieval coverage checks to reduce unsupported responses.

Designed and implemented scalable data pipelines using Python and Apache Airflow for ingesting, cleaning, and processing large-scale retail inventory data from unorganized Kirana stores across India. Applied time-series forecasting and predictive modeling to capture demand patterns under sparse and noisy data conditions. Conducted error analysis and feature evaluation to improve forecast stability and downstream inventory planning decisions. Collaborated with product and engineering teams to deliver a production-grade Universal Inventory Management System (UIM) for client-facing deployment.

### Ph.D. Research Work & Achievements During my Ph.D., I focused on developing advanced AI-driven decision support systems for healthcare emergency management and urban emergency transportation analytics, with a strong emphasis on handling real-world data imbalance, operational constraints, and deployment feasibility. One of my key research contributions was the design and development of **Med-TriageLLM**, a cost-sensitive large language model–based framework for maternal ambulance emergency triage. This work addressed highly imbalanced real-world dispatch datasets where critical cases constituted only about 5% of the total observations. To improve rare-event detection while maintaining operational efficiency, I fine-tuned GPT-3.5 using cost-sensitive prompting strategies and integrated Apriori-derived clinical decision rules during inference. This hybrid reasoning framework significantly reduced false negatives and controlled unnecessary operational alerts. The proposed system achieved a **47% reduction in false positives** compared to a baseline fine-tuned LLM and improved overall classification accuracy to **79.6% on imbalanced evaluation datasets**. It also demonstrated enhanced prioritization capability for rare critical events, with **Lift improving from 1.9 to 2.9** and **F1-score increasing from 15.1% to 21.6%**. Importantly, the framework showed strong **cross-district generalization performance**, reducing false positives by **30–90% compared to zero-shot GPT models** when evaluated on previously unseen regional populations. In parallel, I contributed to a **Department of Science and Technology (Government of India) funded Smart City Emergency Transportation Analytics project (Dec 2019 – Jan 2022)**. In this work, I modeled spatio-temporal emergency medical service (EMS) demand using ensemble learning, clustering methods, and geospatial analytics. The study uncovered significant weather-linked risk patterns and geographic variability in emergency demand, providing actionable insights for resource allocation and urban emergency planning. The findings from this research were later published in **Scientific Reports (2025)**. My doctoral research involved extensive use of **Python, Scikit-learn, time-series modeling, geospatial analysis, GPT-3.5 fine-tuning, cost-sensitive prompting, RAG-style structured reasoning, and association rule mining**, contributing both methodological innovations and real-world impact in AI-enabled emergency healthcare systems.

During my M.Tech, I achieved strong academic distinction by securing the highest CGPA of 9.6 in the first year, ranking among the top performers in the program. Driven by a clear research orientation and interest in applied artificial intelligence, I transitioned directly into a Ph.D. to pursue advanced research in AI-driven decision support systems.

During my B.Tech, I built a strong academic foundation in engineering, mathematics, and computational problem-solving. I consistently maintained a high academic performance throughout the program, developing particular interest and proficiency in areas related to data analysis, algorithms, and intelligent systems.