At Growtrics, I own the question-databank platform end to end. Nightly agents find exam-paper sources, crawlers collect papers, and extraction pipelines turn them into structured questions, with quality gates before a human approves the final push.
I replaced manual labeling with an agentic extraction pipeline using LlamaExtract, Gemini, and shared Pydantic v2 data contracts; it has put 50k+ structured Q&A items at >99% precision into the student-facing product. I also built a chain of specialized agents that produced 30k+ level-precise principal concepts mapped to the multi-market taxonomy.
Previously, at GMO-Z.com RUNSYSTEM, I built and deployed multilingual Vietnamese/Japanese OCR on cloud and mobile, improving accuracy from 94% to 98% while increasing serving throughput. At SmartPay JSC, I developed a real-time fraud detection pipeline and shipped an LLM workflow for interpreting merchant contracts and configuring the internal fee-setting system.
