Temoor Tanveer

I'm a Research Engineer at Qatar Computing Research Institute working on ML systems and LLM safety. My research focuses on building robust, efficient systems: from using LLMs as engines for automated program synthesis and optimization, to designing inference-time safety mechanisms that make models more reliable in deployment.

Previously, I graduated from Carnegie Mellon University with a degree in Computer Science, where I also helped build an open-source federated learning testbed that led to three publications.

Email  ·  GitHub  ·  Google Scholar  ·  LinkedIn

Research Interests

I'm interested in:

• Evolutionary LLM algorithms: Using LLMs to evolve solutions for problems that are easy to verify but hard to solve, including algorithmic discovery, heuristic design, and program synthesis
• ML systems research: Designing machine learning systems that are efficient, robust, and deployable
• LLM safety and security: Adversarial robustness, jailbreak defenses, and alignment

Publications

Beyond RTT: An Adversarially Robust Two-Tiered Approach for Residential Proxy Detection

Residential proxies hide malicious traffic behind legitimate home IPs. We show existing RTT-based detection breaks trivially (99% to 8% recall). Using 900GB of real proxy traffic, we introduce CorrTransform, a Transformer that fingerprints proxy architectures via flow-correlation features, achieving >98% detection even under adversarial attacks.

StructTransform: A Scalable Attack Surface for Safety-Aligned LLMs

We demonstrate that simple structure transformations, changing how prompts are formatted, can systematically misalign state-of-the-art LLMs. This work highlights how brittle current safety training can be to input representation changes.

Federated Learning: Bridging Ideal and Realistic Deployments

Why does federated learning work in simulations but struggle in production? We quantified how device heterogeneity, network conditions, and client churn degrade real-world FL, and built RealFL, an open-source testbed for realistic experiments.

Personal Projects

LEVI: LLM-Guided Evolution for the Price of a Cup of Coffee

CVT-MAP-Elites with AST-based behavioral fingerprinting and tiered model mutations. Beats OpenEvolve, ShinkaEvolve, and GEPA on ADRS benchmarks at 1/3rd the budget.

Chain of Conscience: Adaptive Safety Reasoning at Test Time

Decouples safety reasoning from instruction following in LLMs. The model generates an explicit safety reasoning trace before responding. Trained with GRPO, attack success rates drop from 59% to 2% on jailbreaks, generalizing to unseen adversarial techniques.

Teaching

I served as a teaching assistant at CMU for:

• 15-619 Cloud Computing (Graduate)
• 15-251 Great Ideas in Theoretical Computer Science
• 15-122 Principles of Imperative Computation
• 15-112 Fundamentals of Programming (×2)
• 15-110 Principles of Computing

Design and source code from Jon Barron's website