Teaching at CUNY

AI Approaches for Genomic Analysis (BIO/CS 37105-77105) introduces students to the intersection of RNA biology and artificial intelligence. It covers the fundamentals of gene expression and single-cell RNA sequencing, progressing through computational analysis techniques and into machine learning and transformer-based models applied to biological data. The capstone focus is on virtual cell modeling — using AI to predict how cells respond to genetic perturbations without wet-lab experiments.

LLM Foundations and Mechanistic Interpretability (BIO/CS 47102-79006) takes a deeper dive into the inner workings of the AI models introduced in the first course. Students build transformers from scratch, study the mathematics of self-attention, and then pivot to mechanistic interpretability — a research discipline focused on reverse-engineering what neural networks actually compute. Topics progress from toy models and superposition through circuit analysis, sparse autoencoders, and attribution graphs, culminating in real-world applications like model steering and AI safety.

Foundations of AI Alignment and Safety (CSCI 490) is a project-centered course that builds practical and theoretical competencies in AI alignment research engineering. Students progress through four units over twelve weeks: deep learning fundamentals (implementing CNNs, ResNets, VAEs, and GANs from scratch in PyTorch); transformer mechanistic interpretability (building GPT-2 style models end-to-end, then applying TransformerLens, sparse autoencoders, induction head analysis, and circuit-level techniques such as indirect object identification); reinforcement learning and RLHF (training DQN, policy gradient, and PPO agents, culminating in a complete RLHF pipeline to align a language model to human preferences); and LLM evaluation and alignment science (threat modeling, synthetic dataset generation, and running safety evaluations with the UK AISI Inspect framework). The capstone focus is on analyzing alignment failure modes — including emergent misalignment, deceptive alignment, sandbagging, and sycophancy — grounded in both experimental results and the theoretical literature on AI safety.