About
I am an Engineer & Researcher specializing in LLMs, Computer Vision, and Robotics, focused on building scalable intelligent systems and driving applied AI advancements.
I enjoy exploring the boundary between foundational AI research and its practical applications. Whether it's developing robust agentic architectures or implementing novel computer vision models, I'm driven by the potential to build systems that solve real-world problems.
Publications
APEX-SWE: Benchmarking Software Engineering Agents
Foundational research and evaluation framework benchmarking autonomous AI agents on real-world, complex software engineering tasks.
APEX-Agents
State-of-the-art global evaluation benchmark and leaderboard testing agentic performance on viable, real-world economic workflows.
ACE: The AI Consumer Index
A comprehensive benchmark and evaluation framework designed to assess whether frontier AI models can successfully execute everyday consumer workflows across shopping, food, gaming, and DIY.
APEX-v1: Foundational Agentic Benchmarking
Foundational agentic benchmarking architecture for evaluating Large Language Models on complex, multi-turn software engineering tasks.
CoALM: Conversational Agentic Language Model
Research on unified language models for multi-turn conversations and tool use. Accepted to ACL 2025, exploring advanced conversational AI capabilities.
RWKV Architecture Evolution
Comprehensive review of Receptance Weighted Key Value (RWKV) architecture advancements in efficient language modeling.
Co-Design of Communication and Machine Inference
Cloud robotics research on optimizing communication and inference for networked robotic perception. Published in Robotics: Science and Systems 2021 and Autonomous Robots 2023.
TD-EVAL: Revisiting Task-Oriented Dialogue Evaluation
Novel evaluation framework for task-oriented dialogue systems combining turn-level precision with dialogue-level comparisons.
Occlusion-Aware Tracking for Drones Using Neural Methods
Master's thesis addressing gate tracking under occlusion conditions using synthetic data generation and specialized neural architectures. ViT implementation achieved 99% success rate on occluded points with 82% lower error metrics compared to baselines.