✨ Coming Soon: Adversarial RL for LLM, Specultive Decoding, Triton Kernels, and more ...

🚧 What’s Next: Current Directions and Upcoming Projects

This page offers a concise roadmap of the key research areas and projects I’m actively working on or preparing to launch. These efforts range from foundational LLM optimization to new frontiers in embodied and agentic AI.

⚔️ Adversarial Reinforcement Learning for LLM Alignment

Exploring how adversarial reinforcement learning can enhance the robustness of LLM alignment. The goal is to move beyond static fine-tuning by simulating challenging interactions that force the model to generalize alignment behavior.

⚡ Speculative Decoding for Faster Inference

Implementing speculative decoding techniques to accelerate LLM inference while maintaining output fidelity. This work supports real-time applications and makes large models more accessible in latency-sensitive environments.

🔧 Experimenting with Triton for LLM Speedup

Using Triton to design custom GPU kernels tailored for LLM workloads. The focus is on optimizing memory access patterns and improving throughput during both inference and training.

🧠 Testing Agentic Frameworks for Multi-Step Coordination

Prototyping multi-agent systems built on top of agentic frameworks to coordinate tool use and multi-step reasoning. Emphasis is on modular orchestration, context sharing, and task decomposition.

🦾 Moving Toward Vision-Language-Action Models

Shifting toward embodied AI by working with Vision-Language-Action models that perceive, reason, and act. Also contributing to the development of new benchmarks that better reflect grounded, sequential decision-making tasks.

Get in touch if you want to build together or even just have a chat!