AI Researcher | Language Models & AI Systems
Studying how architectural and capacity constraints shape reasoning and representations in neural systems.

I am an AI researcher working on language models, transformer architectures, and training dynamics, with a focus on architectural analysis, model compression, and interpretability. My work follows an architecture-first approach: implementing and instrumenting models from first principles to study how representations evolve under structural and capacity constraints.

My current research centers on compression-based analysis of large language models, treating pruning, quantization, and distillation as experimental probes rather than purely engineering optimizations.

• LLM Architecture & Compression Research
  Designing and training compact transformer language models to study information bottlenecks, robustness, and representation stability under compression.

• Training Dynamics & Representation Analysis
  Analyzing attention patterns, residual stream behavior, and convergence dynamics across architectural variants and capacity constraints.

• Applied AI Systems (Safety-Critical Contexts)
  Developing agentic retrieval-augmented generation (RAG) systems for complex aerospace engineering documentation, with emphasis on reliability, verification, and structured reasoning.

Status: Active research
Stack: PyTorch, CUDA

Department-led research project studying how architectural choices and compression affect internal representations in transformer language models.

• Training compact transformers from scratch for controlled experimentation
• Applying pruning, quantization, and distillation as analytical interventions
• Instrumenting models to study attention behavior, residual streams, and stability
• Evaluating effects across English and selected Indic languages

Stack: LangChain, LangGraph, Vector Databases

Agentic retrieval systems designed for multi-step reasoning over large, structured technical corpora.

• Query-aware routing between retrieval mechanisms
• Citation-based retrieval to improve reliability
• Focus on structured reasoning in safety-critical domains

• Modeling & Research: PyTorch, CUDA, Transformers
• LLM Systems: LangChain, LangGraph, Hugging Face
• Systems & Tools: Linux, Docker, Git, NVIDIA profiling tools