Projects on Omanshu Thapliyal

Provable Safety and Efficient Memory for Language Models

Mon, 01 Jan 0001 00:00:00 +0000

Problem

Putting language models into safety-critical operations raises questions that accuracy benchmarks do not answer. If a safety filter says a prompt is safe, will it still say so when the prompt is changed slightly? When a model is adapted to a new task, what has the added component learned to remember, and can that be bounded?

Both are questions about dynamics. State space models (SSMs), the recurrent architecture behind Mamba and S4, are dynamical systems, so the tools used to certify controllers apply to them directly: reachability, contraction and model reduction. This project uses those tools to build language-model components whose behavior can be checked, not just measured.

Behind-the-Meter Energy Hub Dispatch for Data Centers

Mon, 01 Jan 0001 00:00:00 +0000

Problem

AI workloads can push a data center’s power demand up by tens of megawatts in minutes, often faster than its grid connection can supply. On-site gas turbines can cover the gap, but they respond over minutes while computing load changes in seconds. Keeping the two in balance without tripping any plant limit is a hard control problem.

Approach

Each side of the facility gets its own learned world model: one for the turbine fleet and one for the AI data hall. A single planner looks ahead across both at once and decides how far the turbines should ramp and how much work the data hall can safely defer. Every command is checked against the turbines’ physical limits before it reaches the plant, so the learned components never act unchecked.

Contract-Aware SLA Compliance Monitoring

Mon, 01 Jan 0001 00:00:00 +0000

Problem

Colocation contracts promise each tenant specific limits on power, temperature and humidity, with penalty credits when those limits are broken. The terms that matter, such as a measurement window, an exclusion or a deadline for claiming credits, are scattered across long contracts and worded differently by every vendor. Monitoring is usually reactive: a breach is noticed only after it has already cost money.

Approach

AI agents read each SLA contract and turn it into structured, machine-checkable rules. Every rule carries the exact contract text it came from, and a verification step checks that quote against the source before the rule is used.

Replicating Expert Behavior via Imitation Learning

Mon, 01 Jan 0001 00:00:00 +0000

Problem

Oil & Gas production assets are run through SCADA systems, where experienced operators read many sensor streams and decide how to act. That know-how is hard to write down as rules. This project set out to learn those decisions directly from recorded expert behavior, so that they can be made autonomously.

Approach

The system is built on Generative Adversarial Imitation Learning (GAIL). Instead of hand-designing a reward, two networks are trained against each other on multi-sensor SCADA data: a discriminator learns to tell the experts' recorded decisions apart from the policy’s, and the policy learns to make decisions the discriminator cannot tell apart from the experts’. The trained policy then reproduces expert behavior.

Safety and Cybersecurity of Cyberphysical Systems

Mon, 01 Jan 0001 00:00:00 +0000

Problem

Drones, air taxis and robot teams are cyberphysical systems: physical machines run by software and linked by networks. Before they can be trusted in safety-critical airspace, we need to know every state they could reach, and whether any of those states is unsafe. The exact answer comes from solving high-dimensional partial differential equations, which is out of reach once a system is nonlinear, learned from data by a neural network, spread across many agents, or under cyberattack.

State Estimation under Communication Uncertainties

Mon, 01 Jan 0001 00:00:00 +0000

Problem

Air traffic surveillance, like many tracking systems, relies on measurements sent over imperfect communication links. Packets get dropped, sometimes at random and sometimes in predictable places, such as regions with poor coverage or radio jamming. A standard Kalman filter assumes every measurement arrives, so its estimates degrade exactly where losses cluster. When many sensors track the same aircraft, each sees only part of the picture and has to agree with its neighbours over a network whose connections keep changing.

Unmanned Aerial Vehicle Safe Autonomous Operations

Mon, 01 Jan 0001 00:00:00 +0000

Problem

Air taxis and delivery drones are expected to fly autonomously over cities, but public trust is low: in one survey cited in this work, only 26% of respondents trusted autonomous air taxis. Before these vehicles can fly on their own, their controllers need a clear safety guarantee: the set of states from which the vehicle can still avoid a collision or reach its landing pad. Computing that set means solving a Hamilton-Jacobi equation, and grid-based solvers become impractical for realistic vehicle models with 6 to 12 states. Learning-based controllers scale better, but they are hard to certify.