Dip Rajeshbhai Shambhvani

A novel benchmarking framework inspired by the ancient art of "Avadhana" to evaluate LLM multitasking, memory retention, and distraction handling.

• Implemented distraction pipelines using LangChain & LangGraph.
• Designed "Memory Recall Score" (MRS) & "Word Overlapping Score" (WOS).
• Benchmarked Llama 3, Mistral, and GPT-OSS under high cognitive load.

A high-performance Redis-inspired database engineered in C++ utilizing asynchronous I/O for massive concurrency.

• Achieved 159k+ GET ops/sec and 133k+ SET ops/sec.
• Implemented RESP-2 protocol and kqueue for non-blocking I/O.
• Scaled to handle 1M+ requests with 1,000 concurrent clients.

A volatile in-memory file system designed for thread safety and low-latency batch operations.

• Engineered mutex-based synchronization for thread safety.
• Achieved ~60µs latency for create/read operations.
• Implemented thread pooling to maximize CPU throughput.

A ViT-GPT2 vision-language model fine-tuned for resilience against image occlusion and corruption.

• Designed ViT-GPT2 encoder-decoder architecture.
• Outperformed SmolVLM baseline in 10-80% occlusion tests.
• Developed a BERT classifier (99.7% F1) to detect generated captions.

A full-stack language processor that simulates register-based hardware to execute custom assembly code.

• Built lexer and parser using Python PLY (Lex & Yacc).
• Simulated register memory architecture and arithmetic logic.
• Implemented complex control flow (branching, loops).

A high-accuracy text classification system for detecting COVID-19 misinformation using BERT variants.

• Fine-tuned TwHIN-BERT achieving state-of-the-art 98.7% accuracy.
• Optimized hyperparameters using Optuna.
• Processed 10k+ tweet dataset with custom tokenization.