At Product Startup, I designed an LLM-assisted document-processing workflow that separates semantic interpretation and routing from deterministic document parsing and assembly.
I developed a multi-stage validation process combining scripted checks, agent-based arbitration, and visual review to identify defects and trace recurring failure classes. I also introduced invariant-based test expansion and regression checks to address defect classes while protecting working scenarios.
I built workflows for generating varied synthetic inputs and edge cases, expanding evaluation coverage where representative real-world data was limited. My work also covered business-process analysis, client communication, UX concepts, and AI-enabled CRM and messaging workflows.
On the Multi-Agent Evaluation & Document Synthesis System, I designed a workflow that processes unstructured input with specialized LLM-based interpretation and produces structured outputs subject to deterministic validation. I organized agent responsibilities and state transitions, and worked on asynchronous task execution, rate-limit handling, and automated output evaluation.
