I'm Thillai,
an ML Engineer
building scalable AI
systems.
Learn More ↓
About Me
About Me
I'm an ML Engineer specializing in LLM systems, scalable inference, and production-scale GenAI infrastructure. Currently building low-latency ML systems for quantitative trading at Summit Securities Group, New York.
My work centers on optimizing LLM inference pipelines - from FlashAttention and speculative decoding to KV cache optimization and CUDA kernel profiling. I build systems that serve models faster and more efficiently at scale.
I'm an active open source contributor to vLLM, llm-d, and EasyEdit, working on the infrastructure that powers LLM serving for thousands of developers. Previously shipped production AI systems at Zideas LLC and conducted research at ISRO and IIT.
What I bring to the table:
Experience
Work Experience
Software Engineer (Machine Learning)
Summit Securities Group - New York, USA
- Building low-latency ML systems for quantitative trading.
Research Assistant
Stony Brook University - New York, USA
- Researching LLM knowledge editing methods and unsafe compliance behavior - investigating how models produce unsafe content in response to unsafe requests through measurable evidence from the pretraining corpus.
- Developing evaluation frameworks to trace model safety failures back to pretraining data, enabling targeted interventions for improving LLM alignment and safety.
Applied AI Engineer Intern
Zideas LLC - New York, USA
- Built a production-grade LLM document intelligence system to autonomously crawl, parse, and validate KYC artifacts across multiple regulatory sources.
- Designed an agentic hybrid RAG + vector indexing architecture with optimized LLM inference via prompt compression and caching.
Computer Vision Researcher
ISRO - Liquid Propulsion Systems Centre, Bengaluru
- Developed a visual defect detection pipeline for X-ray radiography analysis of welded aerospace components using deep learning.
- Designed a SegFormer-based segmentation model integrated with Kubeflow pipelines for automated quality inspection workflows.
Research Intern
Indian Institute of Technology, Tirupati
- Implemented a UniFormer transformer model for liver lesion diagnosis from multi-phase MRI scans.
- Ranked among the top 15 teams globally in the MICCAI Liver Lesion Diagnosis Challenge.
Machine Learning Engineer
BillOK
- Built an OCR model integrated with a language model to process invoices and extract essential fields for financial operations.
- Implemented an automation pipeline linking the system with WhatsApp and email for large-scale invoice processing.
Open Source
Contributions
vLLM
vllm-project
A high-throughput and memory-efficient inference and serving engine for large language models. Contributed to core infrastructure, improving serving performance and developer experience.
View on GitHub ↗
llm-d
llm-d
Distributed LLM serving infrastructure designed for Kubernetes-native deployments. Contributed to the disaggregated serving architecture and deployment tooling for scalable LLM inference.
View on GitHub ↗
EasyEdit (ACL 2024)
zjunlp
An easy-to-use knowledge editing framework for large language models. Contributed to improving model editing capabilities and extending the framework's support for new editing methods.
View on GitHub ↗Education
Academic Background
M.S. in Data Science
Stony Brook University
Graduated: May 2026
B.Tech. in Computer Science and Engineering
Vellore Institute of Technology
Graduated: May 2024
Competencies
Technical Skills
Languages
ML & Inference
Cloud, DevOps & Agents
Highlights
Featured Highlights
Harvard Project for Asian and International Relations
Delegate for HPAIR Asia Conference 2022
Selected as a delegate for the prestigious HPAIR Asia Conference 2022 in New Delhi, presenting on AI solutions for global crises and climate change.
Research Paper - IEEE
Deep Learning-driven Detection of Nuclear Fusion Ignition
Investigated three deep learning architectures - Transformers, LSTM, and ResNet50 - for nuclear fusion event detection. Transformers achieved the highest accuracy.
Read Paper ↗
Research Paper - IEEE
Martian Terrain Classification through Federated Learning
Developed a novel federated learning approach for multi-class Martian terrain classification using DenseNet-121 architecture while preserving data privacy.
Read Paper ↗
Association for Computing Machinery (ACM)
Research and Development Head of ACM-VIT Chapter
Served as R&D Head in 2023, fostering a research-oriented culture through Data Science workshops and mentoring aspiring researchers.
Review Article - MDPI
Exploring Huntington's Disease Diagnosis via AI Models
Comprehensive review of AI-powered algorithms for Huntington's Disease diagnosis, analyzing clinical, genetic, and neuroimaging data.
Read Paper ↗Portfolio
Latest Projects
Multi-Agent AI
Agentic Research Assistant
Multi-agent AI system that automates academic research, literature review, and research paper generation using advanced LLM agents.
View Project ↗
Vision-Language Models
Vision Language Driving Perception
VLM fine-tuning pipeline for autonomous driving with distributed training, TensorRT optimization, and custom evaluation metrics.
View Project ↗
Mental Health AI
CBT-Copilot
Fine-tuned Llama-3.2-3B-Instruct for compassionate CBT-style therapeutic conversations while maintaining professional boundaries.
View Project ↗
Generative AI
Flash AI Search Engine
AI-powered search engine using Gemini 2.0 Flash with live web search results for fast, precise, source-backed answers.
View Project ↗
Generative AI
Dynamic Benchmarking Framework
Dynamic benchmarking framework evaluating LLM accuracy using real-time, location-specific data from WeatherAPI.
View Project ↗
Astroinformatics
Continual LIGO Glitch Detection
Continual learning architecture for LIGO glitch detection using Vision Transformer, achieving 93.4% accuracy in glitch classification.
View Project ↗
Generative AI
MediQuill LLM
Fine-tuned Llama-2 7B on curated medical Q&A data for accurate diagnoses, treatment recommendations, and drug information.
View Project ↗
Astroinformatics
Super Resolution Astronomical Denoiser
SRGAN for galaxy image denoising, improving PSNR by 32.7% and SSIM by 19.8% using transfer learning techniques.
View Project ↗
Fitness Analytics
AI-powered Virtual Fitness Trainer
Real-time exercise tracking using Mediapipe for body landmark detection, angle calculation, and form correction feedback.
View Project ↗