AI Firewall and guardrails for LLM-based Elixir applications

Four main takeaways: (1) LLMs are subject to pressure, they comply despite expressing distress; (2) LLMs are vulnerable to gradual boundary/value violations; (3) when LLMs refuse, they may ignore the response format requirements, so the query is retried; (4) we hypothesise there is a token pattern continuation attractor that might cause obedience.

Train a PI0 flow-matching VLA to weave between obstacles with collision-avoidance baked into the weights via a differentiable safety loss, not a runtime filter.

Safe RL for Autonomous Driving: A project using Constrained MDPs to train a vehicle agent for lane-centering, speed adherence, and collision avoidance. Features kinematic modeling, reward shaping, and safety constraints.