AI for Research Proposal Automation
From 112 research PDFs to a proposal draft—and a human review that caught unsupported citations, inconsistent screening, and an error in the AI judge’s critique.
I supervise metadata specialists at Skydance and work hands-on in streaming quality control for Disney, applying that production judgment to retrieval, triage, and review.
I work at both levels of media operations: supervising specialists, priorities, and escalations while staying close to the assets and decisions themselves. That combination shapes how I design AI-assisted workflows: they need to help the people doing the work, produce results that can be reviewed, and leave important decisions with humans.
From 112 research PDFs to a proposal draft—and a human review that caught unsupported citations, inconsistent screening, and an error in the AI judge’s critique.
Developed as a Johns Hopkins academic project, this inbox-triage system classifies, summarizes, and drafts safe first responses—then evaluates its own outputs against the source emails.
An unimplemented design study for a grounded assistant, built around retrieval before generation, visible sources, explicit uncertainty, fail-closed behavior, and accountable human review.
Make assistance useful, keep its evidence inspectable, and reserve consequential decisions for people. My Johns Hopkins Applied Generative AI certificate projects show both what ran and where it fell short. I also use OpenClaw for working assistants and bounded prototypes.