Phase Adaptive Generation for dLLMs
Academic / Personal
Co-architected Phase-Adaptive Generation (PAG), an ML-driven inference scheduler for Diffusion Language Models (dLLMs) that intercepts LLaDA's decoding loop to dynamically predict block-level control tuples via a dual-head Transformer encoder; reduced average total Number of Function Evaluations (NFE) by 21.4% across 200 GSM8K reasoning tasks while maintaining baseline answer accuracy at 89.5% with negligible predictor latency (~64ms).
LLMDiffusion LLMInference SchedulerPyTorch
Environmental Conditioned Protein Sequence Generation
Academic / Personal
Engineered an environmentally-conditioned protein sequence generator by integrating biological sampling metadata from GOLD and UniProt with NASA POWER and WorldClim geospatial datasets; trained an LSTM architecture to generate functionally targeted amino-acid sequences (e.g., type I antifreeze-like motifs) based on conditional inputs of temperature, precipitation, and solar radiation.
Protein DesignDeep LearningGenerative Models
Large Language Model and Reasoning
Academic / Personal
Built a custom LLM with FlashAttention for faster forward passes and lower memory use, then added post-training alignment pipelines (SFT and GRPO) to improve reasoning quality.
LLMFlashAttentionSFTGRPOPyTorch
Single-Cell Genomics AI Pipeline
Academic / Personal
Developed an end-to-end scRNA-seq pipeline for integration and cell-type annotation using MLPs, autoencoders, and probabilistic models such as scVI and scANVI.
AI/MLGenomicsscRNA-seqscVIscANVI
Distributed Consensus (Paxos)
Personal
Implemented Paxos from scratch in C++, including leader election, log replication, and fault-tolerant state agreement across distributed nodes.
C++Distributed SystemsPaxosFault Tolerance
OS Kernel Development (Minix 3)
Academic
Extended the Minix 3 microkernel with semaphores and process scheduling policy updates, with a focus on context switching, interrupts, and resource management.
Operating SystemsMinix 3KernelScheduling