2026 X-Energy Team
Measuring How Reliably AI Can Report Project Progress
At A Glance:
Capstone Sponsor: X-Energy
Faculty Advisor: Dr. Onur Seref
Capstone Team Members: Vijay Manivannan, Farhad Naseh, Ethan Ruocco, Holly Stewart, Grant Strickland
Solution Summary: Designed and ran a fifty-trial controlled experiment measuring how accurately an AI agent can report project progress from unstructured communication artifacts, then turned the results into a phased roadmap for X-Energy, with full adoption projected to return about 8,788 hours a year to the business.
The Challenge
At X-Energy, project coordinators and team managers spend significant time manually reporting project status and tracking blockers — time that could be better spent on core work. Each week, progress meetings feed a controls function that interprets and inputs the updates, then plans for upcoming milestones. Artificial intelligence could infer that progress directly from communication artifacts, but the company was concerned that AI-generated outputs would lack accuracy and consistency when synthesizing such unstructured sources.
The Solution
Without access to real X-Energy data, the student team built a synthetic company mirroring X-Energy's key functional areas and wrote a ground-truth progress report as the answer key. Around it they generated a body of synthetic artifacts spanning one simulated week — 34 email chains, 16 Zoom transcripts, 15 Slack channel transcripts, three team task trackers, an employee time log, and an employee directory. Feeding these into an AI agent, the team ran five test types across 50 trials and scored every output against ground truth on mean error, root mean square error, R-squared, and F-ratio.
Planned and actual value returned no variation at all, because those figures appear explicitly in structured documents, but earned value exposed the limits: signal dilution, where the agent skims rather than reads once overloaded, and a subjectivity gap that made some teams easier to interpret than others. Meeting transcripts proved the richest context source, capturing the full discussion, clarification, and decision-making process. The team proposed a three-stage rollout — a single planned-and-actual-value agent writing to the project scheduling system in six months, a multi-agent system for well-defined workflows within a year that weights meeting transcripts at 0.5, Slack channels at 0.3, and email chains at 0.2, and a system for ambiguous workflows at two years — all under employee oversight and assurance controls.
The Impact
Removing manual status reporting returns an estimated 8,788 hours a year to the business: 2,288 hours across eleven project control analysts, 2,600 across one hundred team managers, and 3,900 across 450 technical staff. Valued as a long-term asset at a risk-heavy 15% discount rate, that time carries an estimated net present value of 58,586 hours. The work also showed X-Energy's AI implementation team which artifacts an agent can be trusted to read and where human review still belongs.
2026 X-Energy Team Public Presentation