AI Research Intern

About this role

About the role: - Work on AI research projects at Lexsi Labs as an intern, contributing to cutting-edge machine learning and deep learning initiatives - Collaborate with a research team to develop, test, and refine machine learning models and systems - Participate in code reviews, documentation, and development workflows alongside experienced researchers What you'll work on: - Building and experimenting with transformer-based models, understanding how attention mechanisms, positional encodings, and tokenization work in practice - Working with state-of-the-art architectures including BERT, GPT, LLaMA, T5, MoE, Mamba, and similar frameworks - Implementing machine learning solutions using PyTorch with production-quality code standards - Contributing to research in one or more specialized domains within AI What we're looking for: - Strong Python programming skills with the ability to write clean, modular, and testable code - Solid grasp of machine learning and deep learning fundamentals with hands-on PyTorch experience - Thorough knowledge of transformer architectures and core concepts like attention mechanisms, positional encodings, tokenization, and training objectives - Proficiency with Git workflows, software packaging, documentation practices, CI/CD systems, and collaborative development - Strong communication abilities and experience with peer code review and agile team environments Preferred expertise in at least one of these areas: - Explainability and interpretability: Applied work with XAI methods such as DLB, SHAP, LIME, IG, LRP, DL-Backtrace, or Grad-CAM - Mechanistic interpretability: Circuit analysis, activation patching, and feature visualization techniques - Uncertainty estimation: Bayesian methods, ensemble approaches, or test-time augmentation - Model compression: Quantization and pruning techniques for reducing model size, latency, and memory footprint - LLM alignment: Few-shot and zero-shot prompting, chain-of-thought reasoning, RLHF workflows, reward modelling, and human-in-the-loop fine-tuning - Tabular foundation models: Experience with or improvements to TFMs like Orion, TabPFN, or TabICL - Post-training customization: Full-model fine-tuning, LoRA, adapters, instruction tuning, knowledge distillation, or domain-specific specialization