Mojtaba Abdolmaleki

We study pricing and scheduling mechanisms for agentic AI services operated on capacity-constrained compute pipelines. Inference-time compute is endogenous: more compute improves output quality but delays completion and creates congestion. We characterize optimal mechanisms when customers value quality and dislike delay, showing that higher-value customers may optimally receive more compute and higher-quality outputs while being served later. Numerical experiments show that reverse priority can increase profit relative to first-come-first-served and conventional forward-priority policies. Thus, in compute-constrained AI services, premium need not mean faster. It may mean better but slower.

Many ventures create value only upon delivery, but progress and completion are privately observed by the agent running the project. We study a model with moral hazard and limited liability: working generates stochastic breakthroughs while shirking yields private benefits. The agent may exaggerate or delay reporting progress and may strategically sit on completion. We characterize the optimal dynamic contract in this setting. Absent inspections, it delivers truthful reporting via a soft deadline, a probation phase with increasing termination risk during silence, and a short deadline after progress is reported. With inspections, oversight combines random audits during silence to incentivize reporting of breakthroughs, a report-triggered check that the agent is not sitting on a completed project, and late-stage inspections to verify milestone claims.

We develop a homomorphic encryption-based approach to privacy preservation in a dynamic personalized pricing setting. The firm encrypts incoming and historical customer data, estimates demand, and personalizes prices directly on encrypted data without decrypting them. The approach protects all customers, both incoming and historical, and provides perfect privacy protection with no expected-revenue loss, though it is computationally expensive. We also develop a hybrid privacy-preservation approach that reduces computational cost without significant compromise in expected revenue.

We develop a near-optimal dynamic pricing and empty vehicle relocation mechanism for a ride-sourcing system with limited customer patience. The system is modeled as a network of double-ended queues. We establish a fluid limit in a large market regime, show that the fluid-based solution provides an upper bound on performance, and develop a simple dynamic policy based on the fluid solution that nearly achieves that bound. The policy balances supply utilization and customer waiting times under the Square Root Safety staffing rule and is evaluated using empirical data from DiDi Chuxing.

Verifying the provenance of content is crucial for educational institutions, social media platforms, firms, and other organizations. As LLM-generated text becomes almost indistinguishable from human-generated content, we design zero-shot statistical tests to distinguish text generated by different known sets of LLMs and to identify whether text was generated by a known LLM or an unknown source. We prove that Type I and Type II errors decrease exponentially with text length, extend the theory to black-box access via sampling, and validate the method numerically under adversarial post-editing.