<p>Shipped a 16-tool agentic AI support agent into a live ride-hailing platform, building the tool-dispatch and agent orchestration layer from scratch in C# rather than adopting a third-party framework, and owning tool and function calling, tool schema design, argument validation, malformed-output handling, and loop termination.</p><p> Enforced authorization at the tool boundary rather than in the prompt, giving every tool an independent security check so no prompt phrasing could induce the model to return another user’s trip history or financial records, while applying LLM application security and guardrails to the central safety requirement for an agent with live database access in a financial context.</p><p> Made the agent operator-configurable rather than hardcoded through prompt engineering and administrative controls: model version, system prompts, and the human-escalation pipeline are tunable from an admin dashboard without a redeploy.</p><p> Built a Retrieval-Augmented Generation (RAG) assistant over a commercial vendor’s closed-source proprietary language and framework, a corpus with zero public training data where retrieval was the only viable approach.</p><p> Implemented the full Python pipeline end to end: document parsing, chunking strategy, embedding generation, vector storage and similarity search, and context assembly, running on self-hosted Ollama inference to meet the client’s confidentiality constraints.</p><p> Productionized a generative AI dubbing pipeline in Python for a commercial e-learning platform, combining neural speech synthesis for voice cloning with computer-vision facial landmark tracking for lip synchronization to deliver the same course video in 7 languages in the instructor’s own voice.</p><p> Architected the production systems these AI capabilities shipped into: layered ASP.NET Core Web APIs backed by Entity Framework Core with JWT authentication and role-based access control (RBAC), React and Angular front ends in TypeScript, spanning ERP, HR, clinic management, school administration, e-learning, and ride-hailing for 10+ client organizations.</p><p> Diagnosed a critical production failure in which drivers silently stopped receiving ride requests after two or three trips until the app was restarted, with no error and no log entry.</p><p> Traced it across the Flutter client and the SignalR backend hub, outside the primary backend stack, to an undetected real-time connection drop, then led the mobile team through the fix rather than handing off a bug report.</p><p> Owned deployment end to end: authored Dockerfiles and built GitHub Actions CI/CD pipelines from scratch for lint, build, containerization, and deploy; shipped to Azure for some clients and Linux VPS for others; managed Azure SQL with automated EF Core migration runners and Blob Storage with a self-hosted MinIO fallback.</p>
