Member of the Technical Staff

Poetiq, Los Altos

What you will do

Improve the RSI loop – design and build the core algorithms that drive system-level self-improvement.
Push reasoning strategy – propose and test novel, task-specific reasoning strategies; probe LLMs to surface latent knowledge; build data-efficient methods for finding optimal strategies per task.
Build dynamic harnesses – synthesize task-specific execution environments on the fly, even as the system concurrently rewrites its own codebase.
Architect orchestration – design clean abstractions for highly parallelized workflows, turning fragmented signal from thousands of parallel, asynchronous LLM interactions into coherent, high-performance reasoning chains.
Probe capability boundaries – build diagnostic pipelines that separate genuine knowledge gaps from tooling limits (e.g., live retrieval) and reasoning/composition failures.
Close the feedback loop – route knowledge-gap failures into SFT data pipelines; turn reasoning/composition failures into verifiable, exact-match RL tasks.

Poetiq, Los Altos

Improve the RSI loop – design and build the core algorithms that drive system-level self-improvement.
Push reasoning strategy – propose and test novel, task-specific reasoning strategies; probe LLMs to surface latent knowledge; build data-efficient methods for finding optimal strategies per task.
Build dynamic harnesses – synthesize task-specific execution environments on the fly, even as the system concurrently rewrites its own codebase.
Architect orchestration – design clean abstractions for highly parallelized workflows, turning fragmented signal from thousands of parallel, asynchronous LLM interactions into coherent, high-performance reasoning chains.
Probe capability boundaries – build diagnostic pipelines that separate genuine knowledge gaps from tooling limits (e.g., live retrieval) and reasoning/composition failures.
Close the feedback loop – route knowledge-gap failures into SFT data pipelines; turn reasoning/composition failures into verifiable, exact-match RL tasks.