Venture teams evaluating hundreds of startup applications face a bottleneck. Thorough due diligence on any single company takes hours, and even a strong reviewer can only go so deep before moving on to the next application.
In this talk, I'll share my work as an intern on The Open Accelerator team, where I built a multi-agent research and evaluation pipeline using LangGraph to screen the accelerator's Residency applicants. The system fans out subagents to research each startup, then red-teams the findings to catch overstated claims before they reach human reviewers, with clear observability into how each agent reached its conclusions.
I'll walk through the architecture consisting of an orchestrator, research subagents, and a red-team loop. Along the way, I ran into a design question harder than the engineering itself. Figuring out which parts of diligence AI can support well, and which should stay in human hands, turned out to be the real challenge. Attendees will leave with lessons on scoping agentic systems around what AI is actually good at and where to draw the line.