Harshit Purohit

• Designed a novel attentive byte-pair encoding (ABPE) architecture for knowledge graph embeddings, replacing the default tokenizer with a custom-trained one and adding a trainable attention layer, evaluated across 10 KG benchmarks (FB15k-237, WN18RR, YAGO, NELL-995, etc.) on an NVIDIA H100 and university HPC cluster, with comparisons against SOTA baselines like LP-BERT and ULTRA • Extended a retrieval system with 6 decoder-based LLMs (Llama-3, Qwen3) by reengineering their internal attention mechanisms in

• Maintained and optimized open-source R packages (smoof, ecr2) used in evolutionary computation research • Documented the functions and usage of these libraries to support reproducibility and onboarding • Conducted research analysis in evolutionary computation and multi-objective optimization to inform package improvements

• Analyzed company data to identify patterns and provide actionable business recommendations to stakeholders • Used Python to clean and integrate data into the company's database for future use • Collaborated with a team to deliver analysis reports based on the findings

• Built a conversational, multi-platform resume chatbot using Google's Dialogflow API and NLP techniques for B2C use cases • Designed intents, entities, training phrases, and contexts to drive the chatbot's conversational logic • Deployed the chatbot across Google Assistant, Facebook Messenger, Slack, and Telegram as part of a collaborative team