Aditya Kumar Guru

SuperKalam (YC-23)

AI Research Intern

• Project 1: MCQs Generation Pipeline
• Developed a UPSC question generation and quality check pipeline and library using Vertex AI and a Retrieval-Augmented Generation (RAG) engine.
• Project 2: Question extraction & Content Automation
• Built a library for automated question extraction, solution generation and developed a social media post & video creation dashboard using remotion, Gemini and Nano Banana models to streamline educational content and publishing workflows.

AryaXAI (Lexsi Labs)

AI Research Intern

Project 1

• Investigated methods to improve model interpretability and inference efficiency, contributing to the development of an optimization library.
• Benchmarked multiple inference frameworks and libraries to identify critical performance trade-offs, guiding optimization strategies.
• Worked on optimizing inference speed and memory utilization to minimize computational overhead, while assessing scalability across diverse models and datasets.

Artificial Intelligence Institute of South Carolina(UofSC), South Carolina

Research Intern

Project 1: QuickSilver (Under Review)

• Created QuickSilver; implemented multiple FLOP reduction open source libraries, developed layer-wise evaluation pipelines, and implemented token halting and token merging strategies to improve inference efficiency and save computations.

Indian Institute of Science Education & Research (IISER), Kolkata

Research Intern

• Project 3: EACL 2026
• Worked on Logifus; created LogiQAte datasets for Number Series and Direction Sense reasoning, ran SOTA models evaluations, performed mechanistic interpretability using layer-wise next-token log probabilities and contributed to paper writing.
• Project 2: LREC 2026
• Contributed to the development of ObfusQA; conducted observational experiments, LLM benchmarking, intrinsic analysis and layer wise norm-drop analysis, and assisted with paper writing (LaTeX).
• Project 1: NAACL SRW 2025 (Non-archival) , ACL 2025
• Contributed to the development of the SELF-PERCEPT framework through MultiManip dataset creation, dialogue pre-processing, prompt testing, model evaluation and paper writing (LaTeX).

Indian Institute of Technology, Kanpur

Summer Research Intern

Project 1: ReGal (AAAI Bridge 2026)

• Developed ReGal, a PPO based legal reasoning system; engineered PPO training pipelines using Hugging Face TRL library, evaluated reward models, set up evaluation pipelines on Linux based PARAM Siddhi AI server, Vast AI and benchmarked PPO performance on legal judgment prediction and legal summarization across multiple models.