GenAI Optimization |
TOGGLE
Safe LLM compression framework · Python
Temporal-logic-guided compression for edge LLM deployment, focused on balancing efficiency with behavioral constraints.
|
FairCompress
Fairness- and energy-aware LLM compression · Python
Research prototype for compression methods that account for efficiency, fairness, and deployability rather than size alone.
|
VeriNAIS
Signal-temporal-logic-guided neural architecture search · Python
Verified architecture search for AI systems where deployment behavior must satisfy formal constraints.
|
OSPA Transformer
Orthogonal Subspace Projection Attention · Python
Attention mechanism work aimed at improving transformer efficiency through structured subspace projections.
|
AI Hardware Security and Robustness |
flipRL
Reinforcement-learning-driven bit-flip attack research · Python
Prototype connected to efficient bit-flip attacks and accelerator-level vulnerability assessment for LLM systems.
|
VERMITHOR
Runtime orchestration for thermally safe edge inference · Python
Formal runtime control for edge cyber-physical inference where thermal and reliability constraints are first-class requirements.
|
Approximate Computing
Approximate DNN and hardware fault studies · C
Low-level experimentation around approximate computation, fault behavior, and the security-reliability tradeoffs of efficient AI.
|
Neural-Pulse
Runtime verification for LLM behavior · Python
Signal Temporal Logic monitoring for hallucination-like behavior and semantic attack patterns in LLMs.
|
Model Understanding |
GhostTrack
Mechanistic hypothesis and semantic tracking for LLMs · Python
Framework for tracing competing thought trajectories and identifying hallucination risks before final output.
|
Attention Sink Analysis
Transformer attention analysis · Python
Tools for analyzing first-token dominance and attention sink behavior in long-context language models.
|