MS ECE @ Carnegie Mellon University • Researcher • Teaching Assistant • Open-Source Contributor

I'm an MS Electrical & Computer Engineering student at Carnegie Mellon University (Dec 2027) working at the intersection of machine learning, systems, LLM infrastructure, and autonomous systems.

My work spans research, production ML engineering, distributed systems, and open-source contributions.

• 🎓 MS ECE @ Carnegie Mellon University — Dec 2027
• 🧠 Built spatio-temporal graph attention networks for streaming EEG and cortical spreading depolarization detection
• 📄 Research methodology resulted in a patent filing and Society for Neuroscience publication
• 🚗 Evaluated autonomous-driving perception using CARLA + real-world sensor datasets
• ⚙️ Contribute to production ML/LLM infrastructure including bitsandbytes, Qwen3-TTS, and Apple MLX-LM
• 👨‍🏫 TA for Large Language Model Applications and previously Introduction to Deep Learning at CMU
• ☁️ Previously built production ML, LLM, cloud, retrieval, and data systems
• 🏆 CMU TartanHacks Winner
• 💻 420+ LeetCode problems solved

I'm particularly interested in:

ML Systems • LLM Inference • Quantization • GPU Computing • Autonomous Systems • Distributed Systems

University of Pittsburgh | Visiting Researcher

Designed a PyTorch spatio-temporal graph attention network for streaming EEG to detect cortical spreading depolarizations (CSD).

The system models both:

• temporal EEG dynamics
• spatial electrode relationships
• event origin
• propagation behavior
• noisy and constrained signal environments

Outcome: methodology resulted in a patent filing and Society for Neuroscience publication.

PyTorch Graph Neural Networks Attention EEG Signal Processing Time Series

Carnegie Mellon University

Built evaluation pipelines to quantify differences between simulated and real autonomous-driving environments.

Benchmarked CARLA against RaDelft using approximately 1,500 resolution-matched frames per dataset.

Analysis identified:

• up to 3.15× over-representation of dynamic objects
• up to 3.33× under-representation of static classes

The work quantified where simulation distributions diverge from real-world driving data.

CARLA Computer Vision Radar Perception Simulation Python

LLM Quantization / CUDA

🔗 bitsandbytes-foundation/bitsandbytes

PR #1985

Removed an FP16-only bottleneck in the LLM.int8() activation quantization path by adding native BF16 support across Python and CUDA kernels.

Implemented FP32 absmax reduction for improved numerical stability while removing an unnecessary GPU conversion step.

CUDA BF16 FP16 LLM.int8() Quantization PyTorch

Model Training / Debugging

🔗 QwenLM/Qwen3-TTS

PR #336 • Fixes Issue #335

Fixed a Qwen3-TTS 0.6B fine-tuning crash caused by a:

2048 → 1024 embedding dimension mismatch

Restored the missing text-projection path and padding-mask semantics.

Result: converted a training-blocking runtime failure into a successful training step with finite loss = 11.82.

PyTorch Transformers TTS Fine-Tuning Model Debugging

LLM Inference / Speculative Decoding

🔗 ml-explore/mlx-lm

Investigated a Qwen3.6 MTP speculative-decoding failure where multi-turn requests truncated from a requested 400 tokens to as few as 1–9 tokens.

Isolated the failure to MTP-enabled prefix reuse across Apple Silicon systems.

The investigation helped distinguish a model-serving reliability issue from parser or hardware failure.

MLX LLM Inference MTP Speculative Decoding Apple Silicon

Graduate Teaching Assistant | Fall 2026

Supporting coursework covering:

• Transformer-based language models
• LLM applications
• retrieval systems
• agents
• model evaluation
• training and inference infrastructure

Teaching Assistant | Spring 2026

Supported students across:

• neural networks
• optimization
• CNNs
• sequence models
• representation learning
• modern deep-learning architectures

Built a pipeline converting unstructured financial news into structured decision signals using:

• FinBERT sentiment classification
• sentence-transformer embeddings
• named-entity recognition
• temporal feature aggregation
• confidence-calibrated ranking

Combined FinBERT with a lightweight MLP ranking model.

Results

Built a replicated distributed storage system supporting:

• automatic leader election
• replicated-log consensus
• persistent storage
• client request routing
• crash recovery
• network-partition handling
• consistency validation

Python Distributed Systems Replication Consensus Fault Tolerance

Built a Polymarket prediction-market application using blockchain-backed market data.

CMU TartanHacks Winner — February 2026

Built production backend services, REST APIs, and AWS infrastructure across:

EC2 • S3 • Lambda • Docker • CI/CD

Worked on infrastructure supporting a production platform with 3K+ daily active users.

Built a production retrieval and ranking system using:

• Python
• Llama 3
• LangGraph
• CrewAI
• web scraping
• APIs
• structured + unstructured retrieval

Delivered ranked recommendations for 200+ organizations while reducing manual grant-search effort.

LLM Inference • Quantization • BF16 / FP16 • Mixed Precision • GPU Memory • Distributed Training • Model Evaluation

420 problems solved

• 🟢 Easy — 173
• 🟡 Medium — 213
• 🔴 Hard — 34

Interested in ML Systems, LLM Infrastructure, Autonomous Systems, and Applied ML Research.