I build agent systems.
Most of what I care about lives around agents: harness engineering, tool use, post-training, RL-flavored data loops, and the stuff that makes models hold up once they leave the demo stage.
I like working on both sides of the problem: shaping model behavior, and building the engineering around it so that behavior actually survives contact with real tasks.
Agents are just models with tools, memory, and consequences.
Currently spending a lot of time on coding agents, agent workflows, trajectory filtering, evaluation harnesses, and fine-tuning pipelines.
Still learning. Still building. Still trying to make agents a little less fake-smart and a little more useful.
|
An open-source coding agent shaped by harness engineering. Built around agent loops, tool execution control, prompt cache, context compression, and long-horizon task reliability. |
A small-model post-training project built with Qwen3-1.7B + LoRA. Focused on staged SFT, answer-format control, CoT parsing, validation, and eval-driven iteration. |
A multimodal hybrid RAG system for AIOps knowledge bases. Built around dense+sparse retrieval, reranking, heterogeneous documents, and practical retrieval quality. |
