Generative AI Engineer / 生成式AI工程師

★About the role★
The AI Center team at President Information CORP. has not only won the IT Matters Innovation Award but has also published papers in top international journals and obtained numerous domestic and international patents.
You will join our GenAI R&D team to design, build and evaluate agentic and retrieval-based systems. The work is hands-on: you will prototype ideas, run experiments, study how models and agents behave, and turn the findings into clear recommendations.

★What you will do★

• Build and improve LLM agents, including tool calling and multi-step workflows.
• Build and maintain a local agent harness: run agents end to end on your own machine against defined tasks, with tools, sandboxes and model backends that can be swapped in and out.
• Make harness runs reproducible and comparable: fixed task suites, recorded runs and traces that can be replayed, and side-by-side comparison of models, prompts and agent configurations.
• Track agent behavior in the harness (task success, number of steps, tool errors, token use, latency) and investigate regressions and unexpected behavior.
• Design and tune RAG pipelines using advanced retrieval techniques, including graph RAG and integrate vector databases.
• Build document ingestion pipelines that prepare knowledge bases for agents.
• Fine-tune and evaluate models, including LLMs and embedding models, and compare them against off-the-shelf baselines.
• Integrate and compare LLMs across providers, weighing quality, cost and latency.
• Run and extend our evaluation work: build datasets, set baselines, and check that results are statistically meaningful.
• Instrument agents with tracing and use traces to debug quality, latency and cost.
• Develop backend services in Python and work with containerized environments.
• Document findings and share them with both technical and non-technical colleagues.
• Deploy and optimize LLMs using advanced inference engines and frameworks, including Triton Inference Server, vLLM, TensorRT-LLM, and SGLang.
• Ensure low-latency and high-throughput model delivery while balancing hardware and GPU utilization.