GitHub LinkedIn

Erfan Esmaeili, Ph.D.

Machine learning researcher and engineer experienced in model training, evaluation, and deployment. Ph.D. in theoretical physics.

See Projects Download Resume Things I Learned This Week Long Reads I Enjoyed

Selected Projects

View all →

LLM Inference Benchmarking

Comparing LLM inference engines (vLLM, TGI) in latency and throughput.

vLLM TGI FastAPI Docker

Code

RAG QA Engine

Deployed a retrieval-augmented QA engine focused on a history topic.

• Database hosted on Milvus Cloud. Two-stage RAG (reranking).
• Local version used SPLADE and open source models. The deployed version used OpenAI API for text embeddings and llm inference.

Milvus OpenAI text encoder Streamlit

Code Demo

QLoRA adaptation

Adapting open LLMs with LoRA and QLoRA for text summarization.

• Multi-GPU training
• Low Rank Adaptation of quantized models

QLoRA Llama

Code

Papers & Preprints

View all →

Text Embedding is Not All You Need: Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps (CVPR 2025)

Using text-encoder information to optimize cross-attention layers in Stable Diffusion. Our training-free attention control method improves text-image alignment.

Project page PDF

Generate What Matters: Steering Diffusion Models for Targeted Data Generation to Improve Classification (ICLR 2026 under review)

Steering diffusion models to generate effective samples that improve ViT classification accuracy in low-data medical domains.

PDF

7 papers in theoretical physics

On electromagnetic gauge fields, general relativity, celestial holography, and p-form symmetries.

View on Google Scholar →

Skills

Machine Learning Methods

Transformer models; Diffusion models
LLM evaluations; prompt engineering
Retrieval augmented generation
Reinforcement learning
Instruction tuning, SFT, RLHF
PEFT/LoRA/QLoRA

Systems & Tools

PyTorch; HF Transformers & Diffusers
FastAPI; Docker; basic Kubernetes
vLLM/TGI for quick serving prototypes
Python (fluent), Bash; learning C++/CUDA

Research & Analysis

Probability and statistics
Mathematical modeling
Strong analytical thinking;
First principles thinking

Contact

Email: efakhabi@purdue.edu