I am an researcher with a focused interest in mechanistic interpretability and scalable deep learning systems, working toward PhD, My primary research investigates Multimodal Sparse Autoencoders (SAEs) as a framework for disentangling and interpreting latent representations across modalities , probing whether the sparse, overcomplete dictionaries learned by SAEs faithfully recover functionally meaningful features in large neural networks. This work sits at the intersection of representation learning and interpretability, with implications for model transparency and safety.
A second thread of my research concerns the foundational mechanics of automatic differentiation, particularly reverse-mode AD , understanding its computational graph semantics, memory-compute trade-offs, and how these constraints propagate to the design of large-scale training systems. I am interested in how rigorous mathematical formalism here translates to real efficiency gains in practice.
More broadly, I design and study distributed training pipelines and inference architectures using frameworks such as Ray and PyTorch to explore how systems-level decisions interact with learning dynamics at scale. I also write about my research and technical explorations in a series of blog posts, which you can find in the Blogs section.
This project evaluates cross-modal alignment and causal intervention techniques inside transformer backbones. We train high-capacity sparse autoencoders over residual stream activations to discover monosemantic spatial directions that govern coordinate mapping and relational reasoning.
How do modern AI frameworks train massive neural networks without collapsing under the computational weight of calculus? Explore primitives, input/weight Jacobians, vectorization, and reverse-mode AD.
Investigating whether Multimodal Sparse Autoencoders can identify, disentangle, and causally steer spatial reasoning features inside vision-language models without collateral semantic changes.
Extends YOLO+CSRNet crowd counting with a CNN+LSTM pipeline to forecast zone-level crowd density 5–10 minutes ahead and trigger proactive congestion alerts in dense pedestrian environments.
ML pipeline using RDKit, PyTorch, and XGBoost to predict Tg, FFV, Density, Tc, and Rg directly from SMILES strings. Ranked 656/2,400+ overall (65th on private leaderboard with XGBoost).
A proof-of-concept system that segments car parts using YOLOv12, classifies damage severity with a fine-tuned ResNet34, and compiles an automated PDF damage inspection report via Streamlit.
Techniques and memory-efficient strategies for adapting large language models (Codex, StarCoder, Code Llama) to specialized domain codebases using LoRA/PEFT on commodity hardware.
Enter the knowledge graph, a transformative framework reshaping how we structure, navigate, and reason over complex interconnected relational data.
A mathematical breakdown of precision clipping, brevity penalties, and n-gram overlap in evaluating machine translation and text generation systems.
Orchestrated distributed model training and data pipelines using Ray with GPU-aware scheduling. Engineered real-time anomaly detection pipelines for high-volume transactions using deep autoencoders, optimized end-to-end inference (reducing latency by 42%), and built production RAG systems with vector databases.
Designed and deployed scalable data pipelines for multimodal datasets. Optimized distributed training workflows using PyTorch Distributed and Ray (reducing training times by 25%), fine-tuned transformers via PEFT/LoRA, and maintained MLOps infrastructure using Kubeflow and Airflow.
Engineered high-performance data processing services in Rust with Tokio for low-latency data flows. Optimized PostgreSQL access layers and applied performance profiling to achieve a 30% reduction in processing latency.