Back to all jobs
A

AI Engineer (內湖瑞光)

advantech

Taipei_Neihu · tw 13h ago

Job description

We are looking for an AI Engineer to build the models and agents that run on the Advantech AI Platform — our shared “AI Factory” for enterprise AI. This is a hands-on implementation role. The AI Engineer defines the platform architecture and standards; you turn them into working, production-grade models and agents that business teams actually use. Two skills sit at the center of this role. First, model fine-tuning: adapting open-weight LLMs to Advantech’s domains — preparing the data, training, evaluating against a baseline, and serving the result. Second, tool calling: designing tools and function schemas and building agents that call them reliably and safely against real enterprise systems. You will spend most of your time writing code, running experiments, and shipping — and knowing which problems call for fine-tuning, which for better prompting or retrieval, and which for a well-designed tool. Beyond models and agents, you bring full-stack engineering skills: you can build the back-end services, APIs, and front-end interfaces that put them in users’ hands. Platform Domain & Scope Model layer (NemoClaw / LLM models): curate training and evaluation datasets with domain experts; fine-tune open-weight models with supervised fine-tuning and parameter-efficient methods (LoRA / QLoRA), and with preference optimization (DPO / RLHF) where it pays off; benchmark fine-tuned models against base models and prompt-only baselines; and deploy them through the NemoClaw model gateway with optimized inference on on-premises GPUs and in the cloud. Agent layer (tools, skills, and orchestration): design tool and function schemas with structured outputs; build MCP servers and reusable skills; implement multi-step agents that plan, choose tools, recover from errors, and hand off to other agents (A2A); and connect them to real systems such as IoTMart, Data Lake, and SFDC/CRM through governed APIs. Every build follows the platform’s non-negotiables: least-privilege access, prompt-injection defense (including through tool results), human approval for high-impact actions, tracing, eval-driven regression testing, and token-cost awareness. Key Responsibilities Model fine-tuning — Fine-tune open-weight LLMs for enterprise tasks using supervised fine-tuning and parameter-efficient methods (LoRA / QLoRA), and apply preference optimization (DPO / RLHF) where it improves results. Training & evaluation data — Curate, clean, label, and synthesize domain datasets together with domain experts; version datasets and keep their quality high. Model evaluation — Build evaluation sets and benchmarks that compare fine-tuned models with base models and prompt-only or RAG baselines, and use the results to decide when fine-tuning is worth it. Model serving & optimization — Package and deploy models through the NemoClaw model gateway, and optimize inference (quantization, batching, serving engines) on on-premises GPUs and in the cloud. Tool & function design — Design clear tool and function schemas with structured outputs, and implement reliable tool-calling logic with input validation, retries, and error handling. MCP skills & agent orchestration — Build MCP servers, reusable skills, and multi-step agent workflows — planning, tool selection, and A2A hand-offs — on the shared platform. Enterprise system integration — Connect agents to real enterprise systems (e.g., IoTMart, Data Lake, SFDC/CRM) through governed APIs with the right permissions. Tool-use safety — Apply least privilege, output validation, and prompt-injection defense to tool-using agents, and require human approval for high-impact or write actions. Testing & observability — Trace model and tool calls, run eval-driven regression tests in CI/CD, and monitor accuracy, latency, and token cost in production. Retrieval & grounding — Build retrieval pipelines where agents need enterprise knowledge, and combine RAG, tools, and fine-tuned models in the right proportions for each use case. Full-stack delivery — Build the back-end services, APIs, and front-end interfaces (e.g., chat UIs and admin consoles) that turn models and agents into applications people can use. Collaboration — Implement the platform architecture set by the AI Architect, and work daily with AI Project Managers and domain AI Business Partners to turn requirements into shipped solutions. Reuse & knowledge sharing — Contribute fine-tuning recipes, eval harnesses, and tool/MCP templates back to the platform, and share what works with the wider team. Required Qualifications 3+ years of software-engineering experience, including at least 2 years building LLM-based applications hands-on. Hands-on model fine-tuning experience: has prepared datasets and fine-tuned open-weight LLMs with supervised fine-tuning and parameter-efficient methods (e.g., LoRA / QLoRA) using frameworks such as Hugging Face Transformers, PEFT, or TRL, and evaluated the results against a baseline. Hands-on tool-calling experience: has designed tool/function schemas and built multi-step agents that call real APIs, with structured outputs, validation, and error handling. Experience building evaluation sets and metrics for LLM or agent behavior, and using them to make engineering decisions. Strong Python and solid software-engineering fundamentals — version control, testing, and CI/CD. Full-stack engineering capability across back-end services and APIs and front-end development. Experience deploying and serving models or AI services (containers, GPU inference) in the cloud (Azure and/or AWS) and/or on-premises. Working knowledge of retrieval-augmented generation (RAG) and vector search. Clear communication and the ability to work closely with architects, project managers, and domain experts. Bachelor’s or Master’s degree in Computer Science, Engineering, or a related field, or equivalent practical experience. Professional working proficiency in English and Mandarin Chinese. Preferred Qualifications Experience with preference optimization (DPO / RLHF), distillation, or synthetic data generation. Experience building MCP servers, or agent frameworks and A2A patterns. Inference-optimization experience: quantization and serving engines (e.g., vLLM, TensorRT-LLM), or the NVIDIA AI stack on on-premises GPU infrastructure. Experience with LLM observability and eval tooling, and with token-cost management. Security know-how for tool-using agents: least-privilege access, prompt-injection defense, and human-in-the-loop approvals. Hands-on experience delivering full-stack applications end to end (e.g., a React or Vue front end with a Python or Node.js back end). Knowledge of enterprise systems (CRM/SFDC, ERP, e-commerce) and their APIs. A public track record — e.g., published models, open-source contributions, or project write-ups. What Success Looks Like In this role, success means fine-tuned models that measurably beat their baselines where it matters, and agents that call tools reliably and safely against real enterprise systems — shipped to production, adopted by users, and turned into reusable recipes the rest of the team builds on.

Similar open jobs