Edward Kim

“Narrative video at scale requires orchestrated reasoning, not monolithic models.”

Built an end-to-end system that decomposes story prompts into scene graphs, generates visual assets through coordinated agent pipelines, and assembles coherent video narratives. Each stage — script decomposition, visual grounding, temporal consistency, audio alignment — is handled by specialized agents with shared context.

Impact: Demonstrated that multi-agent orchestration can produce structurally coherent video content that single-pass models cannot, with 3x better narrative consistency in blind evaluation.

“Adaptive learning works when the system understands what you actually know, not what you claim to know.”

Developed an AI-powered tutoring system combining real-time ASR with mathematical reasoning. The system listens to a student's verbal explanation, extracts mathematical intent, identifies conceptual gaps, and generates targeted follow-up problems. LaTeX rendering and speech synthesis create a fluid dialogue loop.

Impact: Reduced average time-to-mastery by 40% on pilot concepts compared to static problem sets. Explored at NVIDIA-associated research lab.

“The next order-of-magnitude improvement in AI deployment is hardware-shaped, not algorithm-shaped.”

Conducted deep technical analysis of transformer inference bottlenecks across GPU, TPU, and custom ASIC architectures. Modeled memory bandwidth constraints, quantization tradeoffs, and speculative decoding strategies. Explored dataflow architectures for attention-heavy workloads.

Impact: Produced internal technical report mapping viable paths from current GPU inference to custom silicon for specific model families. Informed hardware procurement decisions.

“Most business processes are graphs of decisions — agents should execute them, not humans.”

Built an agent orchestration platform that converts semi-structured business workflows into executable DAGs. Each node is an LLM-powered agent with tool access, memory, and escalation logic. Handles document processing, data extraction, approval routing, and exception handling across legacy enterprise systems.

Impact: Deployed internally to automate 3 multi-step workflows that previously required manual coordination across 4+ tools.

“The biggest markets are the ones still running on spreadsheets and phone calls.”

Designed and shipped an MVP for digitizing operations in a traditional industry vertical. Built the full stack from authentication and role-based access to real-time dashboards and automated reporting. Focused on reducing time-to-value for non-technical operators.

Impact: Acquired first paying customers within 6 weeks of launch. Validated product-market fit in a space where incumbents move slowly.