Bernie Miao

About

Edtech/education-focused engineer who took an accessibility-critical LLM/vision feature from concept to production: built an OpenCV-gated whiteboard capture pipeline feeding Gemini Vision for handwriting-to-LaTeX, improving math transcription 80% while cutting inference costs 60%. Also built RAG observability and retrieval fixes that stabilized inconsistent answers, and partnered directly with sales to reshape demos and open a new K-12 revenue pipeline aligned to California Digital Divide grant requirements.

Key Strengths

• Designed gated vision pipeline capturing whiteboard equations only during writing bursts
• Delivered measurable impact: 80% improvement in math-content transcription vs audio-only baseline
• Reduced AI processing costs by 60% via selective frame capture instead of full-video processing
• Strong technical judgment in solution selection (rejected audio-only STEM models and standard OCR; costed full-video Gemini approach)
• Built durable AI evaluation infrastructure (benchmarking scripts became team standard for model evaluation)
• Systematic real-time debugging of LLM workflows using end-to-end instrumentation (query, retrieved chunks, similarity scores, outputs)
• Improved RAG answer consistency by fixing retrieval (chunking + metadata filtering) rather than prompt tweaking
• Created ongoing observability layer enabling faster incident triage and regression detection
• Effective developer adoption driver through hands-on demos tied to individual pain points (MCP became team standard)
• Bridged engineering and sales to uncover demo blockers and open new revenue pipeline (K-12 accessibility/grant-aligned pitch)

Experience

Education

Skills

Languages