Eval-Awareness in LLMs
Investigated whether evaluation awareness is encoded in LLM activations using linear probes and
controlled interventions. Showed that strong benchmark performance largely reflects dataset-specific
artifacts rather than true evaluation awareness.
LLMs
Interpretability
Linear Probes
Spotlight: Multi-Agent AI Security Engineer
Built a multi-agent security system with sandboxed autonomous reproduction, verifiable action
provenance, and human-in-the-loop approval workflows.
Multi-Agent
Security
Sandboxing
When Five Agents Behave Like One
Studied correlated errors in LLM committees. Five-agent committees provide only 1.2–1.8 effective
independent voters, motivating a cheaper calibration-based committee strategy.
Multi-Agent
LLM Evaluation
Calibration
Interpretability in POMDPs
Studying internal representations and hidden-state dynamics of DRQN agents in partially observable
environments.
RL
POMDP
DRQN
Interpretability
TiH-IoT Smart Agriculture
Worked at the intersection of drone programming and computer vision for agricultural purposes:
drone navigation, anomaly detection, and task automation.
Computer Vision
Drones
IoT
ACL 2025
DPO Kernels: A Semantically-Aware, Kernel-Enhanced, and Divergence-Rich Paradigm for Direct Preference Optimization
HuggingFace