Research

Systems that make AI behavior inspectable.

Andrei Vince works on reliable AI agents, production infrastructure, computational design, and applied systems where model outputs must survive contact with real state.

01

ACID-Bench

A deterministic fault-injection benchmark for testing whether state-changing tool agents preserve transactional integrity under partial commits, ghost-success retries, stale reads, and recovery failures.

02

Architectural geometry synthesis

Published work connecting cellular automata, structured geometry, conditioned diffusion models, and automated multimodal dataset generation for computational design.

03

Neevel

A screen-free fall-prevention system combining sensor-equipped insoles with a nightlight-shaped hub for balance routines and caregiver visibility. Second place at the 2026 Stanford Center on Longevity Design Challenge.