About this role

▸about Binance — company background

Binance is a leading global blockchain ecosystem behind the world’s largest cryptocurrency exchange by trading volume and registered users. We are trusted by 300+ million people in 100+ countries for our industry-leading security, user fund transparency, trading engine speed, deep liquidity, and an unmatched portfolio of digital-asset products. Binance offerings range from trading and finance to education, research, payments, institutional services, Web3 features, and more. We leverage the power of digital assets and blockchain to build an inclusive financial ecosystem to advance the freedom of money and improve financial access for people around the world.

Binance is looking for a research-minded engineer to join the AI Infra team — sitting at the intersection of frontier model capabilities and real-world agent deployment. You'll work directly with researchers and engineers to push the boundaries of what AI agents can do: from Agentic RAG and context management to task execution, self-evolving agents, and multi-agent coordination.

This is not a pure engineering role and not a pure research role. It's both. You'll be expected to generate original ideas, run experiments, ship prototypes, and iterate fast based on real user feedback. The best candidate is someone who has already internalized agent tools into their daily workflow and has strong opinions about model behavior.

Responsibilities

• Agentic RAG & Engineering: Design and operate next-generation retrieval pipelines — moving beyond static retrieve-once patterns to adaptive, self-correcting, and multi-hop retrieval workflows; architect Agentic RAG systems with dynamic retrieval control, query decomposition, iterative retrieve-reflect-refine loops, and multi-agent retrieval collaboration

• Frontier Harness: Collaborate deeply with researchers and engineers to define and implement model-capability-driven innovations — including context management, long-term memory, subagent and multi-agent architectures, self-evolving agents, and real-word task execution

• Benchmarking & Evaluation: Propose harness-domain and RAG-domain benchmarks and evaluation methodologies; construct benchmark datasets, define annotation strategies, and systematically measure and improve agent intelligence across domains — including retrieval efficiency, latency, groundedness, and task success rate

• Real-world Feedback Loops: Leverage multi-channel user feedback and real-world task data as primary research signals; design experiments and datasets to continuously improve agent and retrieval performance in production scenarios