Applied Machine Learning Engineer focused on LLM training dynamics, evaluation, and inference efficiency.

I work on the parts of ML systems that usually break first:
training stability, memory bottlenecks, retrieval failures, and evaluation blind spots.

Large Language Models

• Instruction tuning (SFT, DPO) and failure analysis
• Attention behavior, KV caching, and long-context issues
• Dataset quality, collapse modes, and reward hacking

ML Systems & Performance

• PyTorch internals (autograd, checkpointing, mixed precision)
• Distributed training (DDP, FSDP, memory tradeoffs)
• Inference optimization (vLLM-style serving, quantization, latency vs throughput)

Evaluation (Underrated, Critical)

• Task-specific evaluation harnesses
• Regression detection across model versions
• Retrieval quality diagnostics for RAG systems

• Analyzed instruction collapse during fine-tuning on noisy datasets
• Compared SFT vs DPO using task-specific win-rate metrics
• Identified normalization and RoPE scaling effects on training stability

• Designed retrieval pipelines with hybrid search (dense + sparse)
• Evaluated hallucination sources caused by embedding drift and chunking
• Built retrieval-quality metrics beyond end-task accuracy

• Reduced GPU memory usage using gradient checkpointing and mixed precision
• Benchmarked inference latency improvements via KV cache reuse and quantization
• Investigated tradeoffs between throughput, batch size, and response time

Core

• Python, PyTorch (deep internals focus)

LLMs

• Transformers, attention variants, KV caching
• SFT, DPO, reward modeling concepts

Systems

• Distributed training (DDP, FSDP)
• Mixed precision, memory optimization

Inference

• Quantization (INT8 / GPTQ-style)
• Serving and performance tuning

Evaluation

• Custom metrics, regression testing
• Dataset and retrieval diagnostics

Most ML failures are not caused by models being too small.
They come from poor evaluation, unstable training, and unexamined assumptions.

I care about:

• Why something fails
• How to detect it early
• How to fix it without scaling blindly

• Repositories here are focused experiments, not tutorials
• Each project prioritizes analysis, metrics, and failure cases
• I value depth over breadth and signal over noise