Skip to content

About

I'm an AI researcher focused on self-improving agents and small language models. I believe everyone should own their own AI - a small model that runs on your hardware, adapts to you, and gets better the more you use it.

The core thesis: capable agents don't just pattern-match on static data, they improve themselves, and they shouldn't need a data center to do it. Modern hardware already runs good-enough small models; the bottleneck isn't scale, it's architecture and efficiency. I co-authored and co-built SIA (Self Improving AI), an agent that modifies both its task-specific scaffolding and its model weights to get better over time - my current best proof of that thesis.

I build transformer and LLM agent architectures, optimize inference and training pipelines, and scale these systems into products that move business metrics. PyTorch-native, CUDA-optimized.

Research Interests

  • Small language models (SLMs): Efficient architectures, on-device inference, quantization, personalization, and making good-enough models that run on everyday hardware.
  • Self-improving agents: Autonomous self-refinement via harness & weight updates, test-time reinforcement learning, and evolving agent scaffolding.
  • Agent evaluation: Adaptive proxy evaluation, MLE-bench, and benchmarks for agents that build agents (AIE-Bench).
  • LLM systems & inference: Scalable inference pipelines, GPU kernel optimization, and production-ready agent infrastructure.
  • Structured reasoning: Questioning-based knowledge activation in AI research agents, latent knowledge unlocking.
  • Applied AI: End-to-end delivery of intelligent products that create real business outcomes.

Philosophy

I believe the path to AGI runs through small, personal, self-improving models. Everyone should have their own AI with them: a model that fits on their hardware, knows their context, and improves on their own schedule. Current hardware already makes this possible; the architectures just aren't there yet. The goal isn't bigger models. It's good-enough models that you can run, tune, and own.

Current LLMs are powerful pattern matchers, but they're static: trained once, deployed forever. Real intelligence requires systems that can modify their own scaffolding and weights, evaluate their own outputs, and improve autonomously over time. SIA is a step in that direction, and I'm pushing further toward self-improving SLMs anyone can run themselves.

Feel free to reach out: [email protected]