Tomas Somoza

I’m a quantitative model risk analyst with deep experience in independent review of AI and quantitative models across a major CIB, combining stochastic processes and derivatives pricing with hands-on machine-learning model validation. At Crédit Agricole CIB, I act as the primary AI specialist for model governance, assessing architecture, implementation quality, and failure modes of deployed GenAI systems.

I lead independent model reviews for risk-critical workflows, including conceptual soundness, feature engineering assessment, backtesting methodology, and limitations under stress—most notably for an LSTM-based GapRisk prediction model for vanilla CDS and CDS Repack structures. I also help shape institution-wide AI risk practices by contributing to the AI Risk framework and defining best practices for LLM and agent-based system evaluation and production monitoring.

I strengthen defense through AI red teaming, covering vulnerability analysis such as prompt injection and jailbreak strategies, and I consistently translate complex model outputs into actionable recommendations for senior management and risk committees. Outside my day job, I work on mechanistic interpretability research using PyTorch circuit tracing and sparse autoencoders, and I developed a thesis model for market crisis prediction using Sornette’s log-periodic power law approach.