首页
/ Evaluation: Acme AI -- Senior AI Engineer

Evaluation: Acme AI -- Senior AI Engineer

2026-09-06 11:36:17作者:裴麒琰

Date: 2026-04-01 Archetype: AI Platform / LLMOps Engineer Score: 4.2/5 URL: https://jobs.example.com/acme-ai-senior-engineer PDF: output/cv-candidate-acme-ai-2026-04-01.pdf


各字段含义:

| 字段 | 语义 | 下游消费者 |
|---|---|---|
| `Date` | 评估日期(YYYY-MM-DD) | 报告文件名 `reports/{###}-{company-slug}-{YYYY-MM-DD}.md` |
| `Archetype` | Step 0 原型检测(archetype detection)的结论 | 决定 B 区块的证据优先级、E 区块的改写方向、F 区块的 STAR 故事选择 |
| `Score` | 1-5 全局分(`X.X/5` 格式) | 4.5+ 立即投递、4.0-4.4 值得投递、3.5-3.9 视情况、低于 3.5 建议放弃(见 [modes/_shared.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/_shared.md?utm_source=gitcode_repo_files) Scoring System) |
| `URL` | 职位链接 | 溯源、去重(`data/scan-history.tsv` 重发检测) |
| `PDF` | 定制 CV 的 PDF 路径 | [batch/batch-prompt.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/batch/batch-prompt.md?utm_source=gitcode_repo_files) 中的 PDF 门限逻辑:分数达到 `auto_pdf_score_threshold`(默认 3.0)才生成 |

报告头部在 [modes/oferta.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/oferta.md?utm_source=gitcode_repo_files) 的 "Post-evaluation" 一节有完整规范,新版还包含 `**Via:**`(中介渠道)、`**Legitimacy:**`(三档真实性判定)、`**Work Auth:**`(工作许可四态)等字段,且紧跟一个 `## Machine Summary` YAML fence 供下游脚本直接解析而不必读散文——字段 schema 以 [batch/batch-prompt.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/batch/batch-prompt.md?utm_source=gitcode_repo_files) 为唯一权威来源(source of truth)。

## 3. Block A — Role Summary:职位画像表

样本中的 A 区块是一张 7 行画像表:

| Field | Value |
|-------|-------|
| **Archetype** | AI Platform / LLMOps Engineer |
| **Domain** | Platform / Infrastructure |
| **Function** | Build |
| **Seniority** | Senior (IC4-IC5) |
| **Remote** | Full remote (US timezone overlap) |
| **Team size** | ~8 engineers |
| **TL;DR** | Senior AI eng to build and scale LLM infrastructure for enterprise customers |

这七个维度就是职位的"分类学坐标":原型(archetype)、领域、职能类型、资历级别、远程模式、团队规模、一句话摘要。Archetype 的取值不是自由文本,而是来自 [modes/_shared.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/_shared.md?utm_source=gitcode_repo_files) "Archetype Detection" 的六分类表(混合时取最接近的两个):

| Archetype | JD 中的关键信号 |
|-----------|-----------------|
| AI Platform / LLMOps | "observability"、"evals"、"pipelines"、"monitoring"、"reliability" |
| Agentic / Automation | "agent"、"HITL"、"orchestration"、"workflow"、"multi-agent" |
| Technical AI PM | "PRD"、"roadmap"、"discovery"、"stakeholder"、"product manager" |
| AI Solutions Architect | "architecture"、"enterprise"、"integration"、"design"、"systems" |
| AI Forward Deployed | "client-facing"、"deploy"、"prototype"、"fast delivery"、"field" |
| AI Transformation | "change management"、"adoption"、"enablement"、"transformation" |

Acme 这条 JD 命中 "evaluation / observability / pipelines / monitoring" 信号,因此被归为 **AI Platform / LLMOps Engineer**。样本中的 TL;DR 一栏正是 [batch/batch-prompt.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/batch/batch-prompt.md?utm_source=gitcode_repo_files) Step 0 中 "buyer intent" 的体现——"Someone who can run AI systems in production with metrics"。

在完整规格 [modes/oferta.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/oferta.md?utm_source=gitcode_repo_files) 中,Block A 还额外要求两个检查项:**Culture screen**(文化筛查,pass / caution / fail,必须写明证据)和 **Work-authorization check**(将候选人的 `config/profile.yml` → `location.authorized_in` 与 JD 的担保声明比对,分为 Sponsors / Not needed / Unstated / No sponsorship 四档;其中 ⛔ No sponsorship 是唯一会压低 location 分并记入 `hard_stops` 的硬阻断)。此外还有一个 **Geo-mismatch check**:职位页面的结构化 location 字段与 JD 正文的到岗要求矛盾时(例如字段写 remote 但正文要求 "hybrid, 3 days in office"),必须在 Block B 顶部追加一条引用 JD 原文的警告行。这些是 sample-report.md 时代规范之上的增量,体现了该格式向后兼容地演进。

## 4. Block B — CV Match:需求→证据映射表 + Gaps

B 区块是整份报告的核心:**JD 需求与候选人简历证据的一行一对应映射**。样本内容:

| JD Requirement | CV Match | Source |
|----------------|----------|--------|
| "Production LLM systems" | Built real-time fraud detection + LLM eval toolkit | cv.md: TechFin Corp |
| "Model monitoring and observability" | Drift detection, Grafana dashboards, retraining triggers | cv.md: ML Platform Lead |
| "Python + distributed systems" | Python, Kafka, Kubernetes, Redis | cv.md: Skills |
| "CI/CD for ML" | Reduced deploy from 2 weeks to 4 hours | cv.md: TechFin Corp |

三列的分工非常明确:**Requirement** 引 JD 原文措辞;**CV Match** 写具体证据而非形容词;**Source** 精确到 `cv.md` 中的段落(如 "TechFin Corp"、"ML Platform Lead"、"Skills"),保证每条断言可回溯。这正是 [modes/_shared.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/_shared.md?utm_source=gitcode_repo_files) Global Rules 中 "Cite exact lines from CV when matching" 的执行形态,也是 [batch/batch-prompt.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/batch/batch-prompt.md?utm_source=gitcode_repo_files) 中 "Never emit a second matrix re-enumerating the same requirements — Block B *is* the requirement→evidence mapping" 规则的原因:第二张同义矩阵迟早与第一张失步,而失步就是自相矛盾的产物。

样本中的 **Gaps** 小节把未匹配项按严重度与缓解策略组织:

| Gap | Severity | Mitigation |
|-----|----------|------------|
| "LLM-specific experience" | Medium | LLM Eval Toolkit is direct proof. Frame fraud detection as "production ML → production LLM" progression |
| "Prompt engineering" | Low | Mention eval toolkit's prompt testing capabilities |

当前 [modes/oferta.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/oferta.md?utm_source=gitcode_repo_files) 的 B 区块规格把这张表升级为五列(`Requirement | Importance | Match | JD signal | Evidence / gap`),并引入三条关键机制,值得在读懂样本后了解:

1. **Two-pass rule(两遍生成)**:Pass 1 只读 JD,在**读 `cv.md` 之前**确定每条需求的 Importance;Pass 2 才读候选证据文件填 Match。Importance 衡量的是"该需求在这条发布中的权重",绝不能被"我正好会这个"锚定——生成顺序本身就是防偏差机制。
2. **Importance 五档 + 证据三层**:档为 `critical / high / meaningful / preferred / low_signal`(枚举而非 0-100 自由数值,因为证据支撑不了 101 个可区分等级);层为 `stated`(JD 明文 must-have,必须附逐字引用)、`structural`(从 JD 结构判断,如所在小节、重复出现)、`inferred`(市场经验推断)。**闸门规则**:`inferred` 行永远不能是 `critical`/`high`,也不能贡献 `hard_stops`——因为高估缺失项的代价是"白丢一次本该投的申请",而低估的代价只是"面试准备稍差",代价不对称。
3. **行预算与排序**:至多 12 行;保留全部 `critical`/`high` 行优先于行预算;排序按重要性降序,同档内**未满足项排在满足项之前**。

Match 的取值是 ✅ Strong / ⚠️ Partial / ❌ Missing / ➖ N/A;且 `critical`/`high` 档的缺失行**必须**在 Gaps 中给出具体面试风险描述 + 缓解策略。样本中 "LLM-specific experience / Medium" 一条正是该模式的前身。

## 5. Block C — Level and Strategy:定级与定位策略

样本内容:

> **Detected level:** Senior (IC4)
> **Candidate's natural level:** Senior-Staff boundary
>
> **"Sell senior" plan:** Lead with platform ownership at TechFin ("led 3-person team, built MLOps for 4 teams"). Frame as ready for Staff scope.

C 区块回答两个问题:这条 JD 要的是什么级别?候选人的自然级别与之差多少?[modes/oferta.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/oferta.md?utm_source=gitcode_repo_files) 规格要求三点:① JD 检测级别 vs 候选人对该原型的自然级别;② **"Sell senior without lying"(不撒谎地卖资历)**——按原型给出具体话术,样本中"led 3-person team, built MLOps for 4 teams"就是可直接复用的措辞,把平台所有权经历框定为 Staff 级scope 的准备度;③ **"If they downlevel me"(被降级录用怎么办)**——可接受的话以薪酬公平为前提,争取 6 个月复评与明确晋升标准。

## 6. Block D — Comp and Demand:薪酬与市场信号

样本中的 D 区块是一张带来源的三行数据表:

| Data Point | Value | Source |
|------------|-------|--------|
| Base salary range | $180-220K | Levels.fyi, similar AI infra roles |
| Total comp (with equity) | $250-320K | Glassdoor estimates |
| Demand trend | High -- LLM infra is top-5 most in-demand | LinkedIn job trends |

注意两件事:每一行都带 **Source**(薪酬数据必须可溯源);"Demand trend" 一行把**市场热度**并入薪酬区块,使 D 成为"值不值得投"的经济学证据。这与受限研究预算(Bounded Research Budget)配套——[modes/oferta.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/oferta.md?utm_source=gitcode_repo_files) 规定 D 与 G 区块合计最多 5 次 WebSearch 查询,"This mode is an evaluation workflow, not deep company research",深度公司研究应另跑 `/career-ops deep`。

在现行规格中,D 区块还强制先做**公司类型分类**([modes/_shared.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/_shared.md?utm_source=gitcode_repo_files) "Company Type and Compensation Reliability"):从 Public big tech(薪酬可信度高到中)到 Local SMB(低)共 10 类,因为公开薪酬数字的可信度因公司类型而异;再判定**薪酬可靠性档位**(High/Medium/Low/Unknown):出现 "comprehensive salary"、"up to"、"OTE"、"uncapped" 等措辞时降为低可靠;JD 未报薪时该区块收敛为两行(公司类型 + 可靠性档位),且薪酬数字必须逐字进 Machine Summary 的 `advertised_comp` 字段,供 [salary-gap.mjs](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/salary-gap.mjs?utm_source=gitcode_repo_files) 消费。薪酬维度的打分锚点是:5 = 前四分位,4 = 高于市场,3 = 市场中位,2 = 略低,1 = 明显低于市场。

## 7. Block E — Personalization Plan:简历改写清单

样本内容:

| # | Section | Current | Proposed Change | Why |
|---|---------|---------|-----------------|-----|
| 1 | Summary | "Full-stack AI engineer" | "AI platform engineer focused on LLM infrastructure and observability" | Match JD language |
| 2 | TechFin bullets | Generic ML platform | Add "LLM serving" context | JD specifically mentions LLMs |
| 3 | Projects | Both listed equally | Lead with LLM Eval Toolkit | Direct LLM experience proof |

E 区块是一张**可执行差异表**:定位简历中哪个段落(Section)、现状(Current)、改成什么(Proposed Change)、为什么改(Why)。规格要求覆盖 Top 5 CV 改动 + Top 5 LinkedIn 呈现改动。表中的 Why 列都是"对齐 JD 语言"(Match JD language / JD specifically mentions LLMs / Direct LLM experience proof)——定制的本质是**用 JD 的词汇重述既有的真实经历**,而非虚构技能;这也是 [batch/batch-prompt.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/batch/batch-prompt.md?utm_source=gitcode_repo_files) 中 PDF 生成规则 "Inject keywords ethically into existing achievements; never invent skills or metrics" 的上游依据。

## 8. Block F — Interview Plan:STAR 故事矩阵

样本内容:

| # | JD Requirement | STAR Story | S | T | A | R |
|---|---------------|------------|---|---|---|---|
| 1 | Production LLM systems | FraudShield scaling | 10K TPS requirement | Built streaming pipeline | Kafka + ensemble + feature store | 99.7% precision, $2M saved |
| 2 | Team leadership | ML Platform team | 4 teams needed MLOps | Led 3-eng team, built platform | Registry + A/B + feature store | Deploy time 2 weeks → 4 hours |

> **Recommended case study:** LLM Eval Toolkit -- shows LLM-specific expertise + open source impact

F 区块把 JD 需求逐条映射到 STAR(Situation / Task / Action / Result)结构的故事上,并给出 1 个推荐案例研究。现行规格([modes/oferta.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/oferta.md?utm_source=gitcode_repo_files) 与 [batch/batch-prompt.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/batch/batch-prompt.md?utm_source=gitcode_repo_files))要求 6-10 个 **STAR+R**(多一列 Reflection,"junior candidates describe what happened, senior candidates extract lessons"),故事按原型选帧(LLMOps 强调 metrics/evals/production hardening,Agentic 强调 orchestration/HITL),并同步写入 `interview-prep/story-bank.md` 形成可复用的故事库,外加红旗问题预案。

## 9. 尾部 Keywords Extracted:ATS 关键词池

样本最后一节是从 JD 提取的关键词列表:

> LLM infrastructure, model serving, observability, ML platform, distributed systems, Python, Kubernetes, model monitoring, CI/CD, prompt engineering, evaluation, production ML, enterprise AI, scalability, reliability

规格要求 15-20 个 JD 关键词([modes/oferta.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/modes/oferta.md?utm_source=gitcode_repo_files) 记作 "Keywords extracted",batch 模式记作 "Extracted keywords")。其用途是双重的:ATS 匹配的词表,以及 CV 定制时"ethically inject"(注入现有成就措辞)的素材。PDF 生成的设计约束(单栏、标准节标题、UTF-8 可选文本、关键词自然分布于 summary/experience/skills/projects)见 [batch/batch-prompt.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/batch/batch-prompt.md?utm_source=gitcode_repo_files) Step 4。

## 10. 报告的下游消费:编号、归档、tracker 与前端渲染

**编号与命名。** 报告落盘为 `reports/{###}-{company-slug}-{YYYY-MM-DD}.md`,三位序号必须通过 `node reserve-report-num.mjs` 原子申领(stdout 返回 `{###}`),写完后用 `node reserve-report-num.mjs --release {###}` 释放哨兵,防止并发竞争(对应测试 [tests/reserve-report-num.test.mjs](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/tests/reserve-report-num.test.mjs?utm_source=gitcode_repo_files))。中介职位雇主未知时用 `confidential-{agency-slug}` 命名且**永不**重命名,只改标题与 YAML 头。

**Machine Summary 与 JD 归档。** 完整规格中报告紧跟头部携带 `## Machine Summary` YAML fence(`company`、`role`、`score`、`legitimacy_tier`、`final_decision`、`hard_stops`、`advertised_comp`、`requirement_importance`、`risk_summary` 等键,规则见 [batch/batch-prompt.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/batch/batch-prompt.md?utm_source=gitcode_repo_files)),并强制附 `## Job Description (archived verbatim)` 原文归档段——长 JD 可用 `archive-posting.mjs --report={num}` 外置,由 `check-jd-archive.mjs` 校验,且该校验被 `test-all.mjs` 纳入回归([tests/check-jd-archive…](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/tests/jd-capture.test.mjs?utm_source=gitcode_repo_files) 相关)。[test-fixtures/upgrade/state-v1.18/reports/001-acme-2026-06-20.md](https://gitcode.com/GitHub_Trending/ca/career-ops/blob/afc5275d10ca22749d3d7528167961bb97d95939/test-fixtures/upgrade/state-v1.18/reports/001-acme-2026-06-20.md?utm_source=gitcode_repo_files) 展示了新版头部的最小形态:

```markdown
**Score:** 4.2/5
**Legitimacy:** High Confidence
登录后查看全文
热门项目推荐
相关项目推荐

项目优选

收起
kernelkernel
deepin linux kernel
C
33
18
ops-transformerops-transformer
本项目是CANN提供的transformer类大模型算子库,实现网络在NPU上加速计算。
C++
1.14 K
2.75 K
pytorchpytorch
作为 Ascend for PyTorch 社区的核心组件,TorchNPU 是昇腾专为 PyTorch 打造的深度学习适配插件,使 PyTorch 框架能够直接调用昇腾 NPU,为开发者提供昇腾 AI 处理器的超强算力。
Python
857
1.35 K
docsdocs
暂无描述
Markdown
898
5.82 K
ops-nnops-nn
本项目是CANN提供的神经网络类计算算子库,实现网络在NPU上加速计算。
C++
921
1.84 K
jiuwenswarmjiuwenswarm
JiuwenSwarm 是一款基于openJiuwen开发的智能AI Agent,它能够将大语言模型的强大能力,通过你日常使用的各类通讯应用,直接延伸至你的指尖。
Python
3.8 K
1.02 K
kernelkernel
openEuler内核是openEuler操作系统的核心,既是系统性能与稳定性的基石,也是连接处理器、设备与服务的桥梁。
C
531
596
cann-learning-hubcann-learning-hub
CANN 学习中心仓,支持在线互动运行、边学边练,提供教程、示例与优化方案,一站式助力昇腾开发者快速上手。
Jupyter Notebook
1.02 K
519
ops-mathops-math
本项目是CANN提供的数学类基础计算算子库,实现网络在NPU上加速计算。
C++
1.36 K
1.46 K
AscendNPU-IRAscendNPU-IR
AscendNPU-IR是基于MLIR(Multi-Level Intermediate Representation)构建的,面向昇腾亲和算子编译时使用的中间表示,提供昇腾完备表达能力,通过编译优化提升昇腾AI处理器计算效率,支持通过生态框架使能昇腾AI处理器与深度调优
C++
548
391