Phaestus uses workspace video to build simulation-ready scenes and study whether a robot could perform a specific task. I’m building toward evidence-backed reports that identify likely failures and recommend changes to the robot, tooling, or layout.
A Berkeley Robotics Hackathon 2026 winning project that pairs first-person and surrounding-world views to help robots learn from human demonstrations. On Assembly101, this pairing raised held-out task-labeling accuracy from 8.0% to 14.2%.
A browser simulator that measures how a robot changes pedestrian movement by running the same synthetic crowd with and without the robot. Paired runs provide an answer key for evaluating disturbance metrics and their false alarms.
A Fantasy Premier League decision engine that forecasts player points and compares every legal single transfer with holding the current squad. It keeps manager goals, constraints, and assumptions explicit throughout the recommendation.
A local-first memory layer that syncs personal context across ChatGPT, Claude, Poke, and Claude Code. It extracts durable facts from conversations, flags conflicts, and sends approved memories back in each platform’s format.
A task recommender that uses a project’s development history to suggest what to build next. Its evaluation compares recommendations with the work shown in actual commit diffs.
A voice interview platform for staffing teams to create role-specific phone screens and assess candidates. It turns interview transcripts, workspace activity, and behavioral signals into searchable evaluations.