Immanuel Lewis

Architected and deployed a production-grade, isolated LLM testing environment for Citi’s global fraud department. Engineered an end-to-end Python GenAI backend for intelligent document processing including automated classification, dynamic schema generation, multi-layer validation, and 95%+ extraction accuracy across 60+ heterogeneous fraud document types. Designed context-aware prompt chaining and intelligent PDF segmentation to handle real-world investigation datasets. Built human-in-the-loop review workflows that reduced document review cycle time by 20% while serving as a QA interface between model outputs and fraud analysts. Also designed automation for transaction management and global loan operations, eliminating manual processing of ~29,000 new and amended deals annually by integrating AI validation services into operations workflows.

Executed adversarial red-teaming for cutting-edge conversational AI systems with a focus on prompt injection, jailbreak attempts, input/output evasion, and real-world misuse scenario testing. Conducted systematic adversarial testing campaigns to identify critical exposure points across diverse AI use cases aligned with GenAI attack vectors. Produced actionable insights that improved AI threat detection, monitoring, and model reliability to support safer enterprise deployment of AI technologies.

Built multiple production-grade AI systems for clients across hospitality, professional services, and quantitative research, with emphasis on scalability and performance. Self-developed agentic AI platforms such as an AI interview and legacy intelligence system (“Living Inheritance”) featuring multimodal/voice AI, evidence-grounded LLM workflows, source-linked evidence capture with timestamps, and reviewable memoir outputs. Architected production full-stack orchestration components including consent workflows, media ingestion, idempotent transcription, citation-backed insight generation, photo integration, and multi-stage human review. In development, designed an agentic engineering control plane for multi-repository autonomous engineering with governance, CI/CD integration, auditing, telemetry, release judgment, staging, incident repair, bounded autonomy, evidence-based decision gates, and human approval for high-risk actions.