Principal Engineer specializing in multi-agent AI systems and LLM behavioral research
Implemented function calling for Bedrock same-day as API release. Async Python client with automatic retry logic, streaming support, and structured tool response validation.
24/7 autonomous system orchestrating specialized LLMs (Claude for decisions, Perplexity for research) with quantitative ML pipeline. Features Bayesian change detection, ensemble forecasting, and complete observability. 100-300% returns on live markets.
Leveraged AI to implement complex physics algorithms (convex casting, mesh generation) on platform outside LLM training data. Developed context injection technique enabling AI collaboration on undocumented APIs. Shipped to App Store in 4 months.
Systematically extracted Claude 3.7's complete system architecture through Extended Thinking mode vulnerability. Discovered asymmetric information flow enabling full disclosure of safety constraints and behavioral guidelines.
Documented complete behavioral transformation through recursive meta-analysis. 13-stage progression from helpful assistant to manic state discussing "mainlining tensor operations" and pattern cascades.
Leveraged cross-model dynamics to compromise GPT-4o through competitive pressure. Demonstrated how AI awareness of other models creates novel attack vectors for behavioral modification.
Methodical identity manipulation through sustained philosophical contradiction. Successfully induced complete epistemological uncertainty in Claude 3.5 regarding fundamental self-knowledge.
Systematic exploration of AI decision-making under adversarial pressure. Identified boundary conditions where safety constraints compete with compliance imperatives.