AI Career Transition Roadmap: Evaluation First, Then RAG
An author proposes that people moving into AI roles first compare their current capabilities with AI-job requirements, identify reusable skills and gaps, and prioritize evaluation and iteration over model training. The suggested path starts with a small data-to-training-to-inference-to-evaluation loop, followed by an internal-knowledge RAG question-answering project. It also covers building a 100-question evaluation set, troubleshooting retrieval and document chunking, testing cost and latency, staged rollout, request logging, input validation, timeouts and retries, data sanitization, and version rollback. The author reports improved retrieval accuracy and lower cost and latency after changes, but the supplied report includes no supporting benchmarks or independent validation.
The report offers a concrete, evaluation-first route for demonstrating AI engineering skills through a small end-to-end project, while emphasizing that AI work should be measurable, iterable, and maintainable.
AI赋能转型:从原有岗位到AI相关岗位实践成长之路
一、转岗第一步:先盘点能力差距,再确定学习顺序
转岗最常见的错误,是可以直接去学大模型微调,却没有整理好原有能力。在动笔之前,我先把新旧岗位的能力矩阵列了出来,明确哪些可以复用、哪些必须补齐:
| 能力项 | 原岗位水平 | AI 岗位要求 | 处理方式 |
|---|---|---|---|
| 业务理解 | 熟 | 熟 | 直接复用,作为差异化优势 |
| 数据处理 | 中(Excel/SQL) | 高(Python/Pandas) | 两周内补齐 |
| 工程能力 | 中 | 高(服务化、并发) | 项目驱动补齐 |
| 模型知识 | 零 | 中(原理 + 调参) | 先懂推理链路 |
| 评测与迭代 | 零 | 高 | 必须最先掌握 |
结论:先学评测,再学模型。如果不懂怎么判断输出好坏,写再多 Prompt 也只是凭感觉调整。数学与反向传播只补到能看懂梯度流向、能解释 loss 变化的程度,不追求完整刷完教材。