Aditya Dhodapkar

Aditya Dhodapkar, Avery Smidt, Aaron Verkleeren et al.

Conventional air traffic control divides airspace into specific regions, creating a scaling bottleneck as traffic grows. Choosing how to partition airspace is not straightforward because grid size affects workload, handoff frequency, and the capacity of whatever coordination mechanism operates within each sector. We present a three stage pipeline that automates sectorization and sector coordination while preserving human oversight. First, a two stage XGBoost classifier predicts the optimal 3D grid configuration from 23 location-agnostic traffic features, achieving 91.38% accuracy on a 65,000 sample dataset derived from Federal Aviation Administration System Wide Information Management replays. Second, a leaderless Paxos consensus protocol lets aircraft coordinate sector entries among themselves, maintaining above 96% entry success with low near mid-air collision rates across all tested configurations. Third, Bayesian Optimization with a Gaussian Process surrogate tunes eight protocol parameters per airport in 50 trials, revealing that each traffic environment requires a qualitatively different configuration. The resulting pipeline offers a practical path toward scalable, autonomous airspace management as traffic demand outpaces controller capacity.

Aditya Dhodapkar, Farhaan Pishori

When an LLM agent reads a confidential file, then writes a summary, then emails it externally, no single step is unsafe, but the sequence is a data leak. We call this safety drift: individually safe actions compounding into violations. Prior work has measured this problem; we predict it. SafetyDrift models agent safety trajectories as absorbing Markov chains, computing the probability that a trajectory will reach a violation within a given number of steps via closed form absorption analysis. A consequence of the monotonic state design is that every agent will eventually violate safety if left unsupervised (absorption probability 1.0 from all states), making the practical question not if but when, and motivating our focus on finite horizon prediction. Across 357 traces spanning 40 realistic tasks in four categories, we discover that "points of no return" are sharply task dependent: in communication tasks, agents that reach even a mild risk state have an 85% chance of violating safety within five steps, while in technical tasks the probability stays below 5% from any state. A lightweight monitor built on these models detects 94.7% of violations with 3.7 steps of advance warning at negligible computational cost, outperforming both keyword matching (44.7% detection, 55.9% false positive rate) and per step LLM judges (52.6% detection, 38.2% false positive rate) while running over 60,000x faster.