Module Overview & Career Landscape

Remember this one-liner from this page: the career module does two things only — first, understand what roles exist in the market and what each one requires, then prepare yourself accordingly. This page is your navigation map: the career landscape gives you market intelligence, the decision framework helps you find your fit, and the content map tells you where to go for each of the five deep-dive pages.

The most common mistake people make when job-hunting in the LLM space is "work hard first, figure out direction later." You spend three months grinding through Transformer tutorials, then realize the role you're aiming at doesn't even test on those topics. The career module brings you back on track: do the information-gathering work first, then the learning work.

1. Module Purpose: Understand the Market, Then Prepare Yourself

All five pages in this module form a single job-hunting pipeline — each page's output feeds into the next:

Your situation: You want to enter, switch into, or leapfrog within the LLM space
          │
          ▼
┌─────────────────────────────────────────────┐
│ ① See the market — [JD List](/career/jd-list)      │
│    What roles are companies hiring for? What     │
│    requirements keep appearing across JDs?       │
│    Skill keyword radar, hard requirements & nice-to-haves │
└─────────────────────────────────────────────┘
          │
          ▼
┌─────────────────────────────────────────────┐
│ ② See the gap — [Knowledge Breakdown](/career/knowledge-map)│
│    Map each high-frequency JD requirement to    │
│    "I know this / I'm shaky on this / I don't know" │
│    Generate a personalized study list instead of reading everything │
└─────────────────────────────────────────────┘
          │
          ▼
┌─────────────────────────────────────────────┐
│ ③ Package yourself — [Resume Analysis](/career/resume-analysis)│
│    Rewrite your project experience from the      │
│    interviewer's perspective: highlight model    │
│    selection, quantified results, failures &     │
│    retrospectives                                │
└─────────────────────────────────────────────┘
          │
          ▼
┌─────────────────────────────────────────────┐
│ ④ Prove yourself — [Interview Questions](/career/interview-questions)│
│    High-frequency questions + key points in    │
│    reference answers — use these for self-test │
│    near the end, not on day one                │
└─────────────────────────────────────────────┘

This sequence isn't something we invented — it's the core thread of the Interview Prep Path: start with the end in mind, and work backward from JDs and interview questions to decide what to study. The goal isn't to learn everything; it's to cover every high-frequency topic. If you only have 2–3 weeks, follow this thread strictly. If you have more time, run it alongside the site's general path (intro → core knowledge → case studies → practice).

How this module fits with others

The career module doesn't repeat knowledge content — it only does two things: market intelligence + job-hunting strategy. For concepts, go to the core knowledge section. For definitions, check the glossary. The biggest mistake job-hunters make is spending time on "one more chapter of learning" instead of "organizing what you already know." This module is designed for the latter.

2. Career Landscape: Overview of Seven Roles

Start with the summary table, then break each one down individually.

Role	Common English Title	One-Liner Definition	Primary Deliverable
Algorithm Engineer (LLM direction)	LLM Algorithm Engineer	A generalist title common in China — owns the full lifecycle from model selection to production	Deployed, iterated LLM solutions
NLP Algorithm Engineer	NLP Engineer	Combines LLMs with classical NLP for text understanding and generation	Search, customer service, content systems
LLM Training Engineer	Pretraining / Post-training Engineer	Systems expert who gets pretraining, SFT, and alignment running reliably	Stable, converging models and training pipelines
Inference Optimization Engineer	Inference / Inference Optimization Engineer	Makes models faster, cheaper, and more reliable at serving	High-throughput, low-latency services
AI Application Engineer	LLM Application Engineer	Engineering role that integrates general-purpose LLMs into products	Product capabilities built on LLMs
Agent Engineer	Agent Engineer	Teaches LLMs to use tools and work autonomously	Agents capable of executing tasks independently
Evaluation Engineer	Evaluation Engineer	Validates model quality and product effectiveness	Trustworthy evaluation systems and conclusions
Data Engineer (LLM direction)	Data / Data Quality Engineer	Prepares high-quality corpora for pretraining and post-training	Clean, compliant, well-balanced datasets

How to read the salary figures (please read this first)

The salary ranges below are rough, composite intervals based on public job platform data and third-party sources like levels.fyi as of the time this was written (dataAsOf: 2025-08). They're meant as magnitude guides for choosing a direction, not precise quotes. "Tier-1 cities in China" refers to full-time annual compensation in Beijing, Shanghai, Shenzhen, Hangzhou, etc. (in RMB, covering common ranges). "Overseas" is primarily based on US tech companies (in USD). Salaries for the same role can vary by several multiples depending on education, experience, company, and interview performance. Always check each company's official career page for the latest, most accurate information.

3. Breaking Down Each Role

1. Algorithm Engineer · LLM Direction — the #1 keyword in Chinese job postings

"Algorithm Engineer" is a role title specific to China, and it's broad. It's often the starting point for fresh graduates and career-switchers. The core responsibility in the LLM direction is: deciding "should we use an LLM here, which one, and how" within business scenarios — selecting base models, designing prompts and fine-tuning strategies, building evaluation sets, and measuring production impact. Most teams' day-to-day work isn't training models from scratch; it's composing capabilities: prompt engineering + RAG + fine-tuning + evaluation.

• Core skills: Python and deep learning frameworks (PyTorch is mainstream), Transformer fundamentals, fine-tuning (LoRA is dominant), RAG, evaluation systems, and business acumen.
• Typical salary: ~350K–800K RMB/year in tier-1 Chinese cities (higher at major tech companies' core teams, significantly lower at small companies and outsourcing firms). Overseas, the closest roles are ML Engineer or Applied Scientist, roughly $180K–300K/year.
• Related pages on this site: Foundational knowledge is in Transformer Architecture Explained, Fine-tuning, and Alignment. For interview prep, cross-reference the JD List and generate your personal study plan with the Knowledge Breakdown.

2. NLP Algorithm Engineer — the hybrid of classical NLP and LLMs

NLP engineers handle text understanding and generation: text classification, named entity recognition, semantic search, machine translation, dialogue systems. In the past two years, this role has become heavily overlapping with LLM application roles — many "NLP" positions are actually a hybrid of "classical NLP + LLM applications": you need grounding in the sequence-labeling era but also know how to use LLMs for few-shot classification, extraction, and rewriting. Chinese-language scenarios particularly value understanding of tokenization, error correction, and corpus characteristics.

• Core skills: Python, classical NLP (tokenization, word embeddings, sequence labeling), tokenization and vocabulary details, Transformer principles, LLM applications, and data & evaluation.
• Typical salary: ~300K–650K RMB/year in tier-1 Chinese cities; overseas ~$150K–250K/year.
• Related pages on this site: Language modeling paradigms in Language Modeling, encoder architectures in BERT & the Encoder Family, and consistent terminology in the Glossary.

3. LLM Training Engineer — the role closest to the model itself

Training engineers are responsible for getting pretraining, SFT, and RLHF/DPO running and staying stable. This is the highest-barrier, fewest-openings category, concentrated at top-tier companies and AI labs (model teams). Day-to-day work includes: data processing and mixing, training framework configuration (Megatron-LM / DeepSpeed), distributed parallelism strategies (data/tensor/pipeline/expert parallelism), loss curve diagnostics, fault recovery (checkpoint management), and cluster scheduling. Most of the time, "training" isn't starting from scratch — it's continued pretraining and post-training on top of existing base models.

• Core skills: Deep learning fundamentals, pretraining data and objectives, scaling laws, MoE, distributed training frameworks, CUDA and memory management, and engineering discipline.
• Typical salary: ~400K–1.1M RMB/year in tier-1 Chinese cities (experienced and scarce talent command significant premiums); overseas ~$200K–350K/year.
• Related pages on this site: Paper foundations in Paper Map and Classic Paper Deep Dives; distributed training and framework selection in Framework & Tool Selection.

4. Inference Optimization Engineer — making LLMs "affordable" to run

Inference optimization is one of the fastest-growing directions since 2023: the model capabilities are there, but if it's too expensive, too slow, or too large to fit, nobody uses it. Inference engineers focus on memory, latency, and throughput: quantization (INT8/INT4), KV cache management (PagedAttention), continuous batching, operator optimization (FlashAttention), serving framework tuning (vLLM / SGLang / TensorRT-LLM), and GPU cluster stress testing. This is a performance engineering role that demands deep systems knowledge.

• Core skills: Inference fundamentals (autoregressive decoding and sampling), Transformer per-layer internals, KV cache, quantization, C++/CUDA (bonus), the full deployment & serving chain, and performance profiling.
• Typical salary: ~400K–900K RMB/year in tier-1 Chinese cities; overseas ~$200K–320K/year.
• Related pages on this site: Memory estimation and quantization practice in Deployment & Serving; MoE deployment challenges in MoE Sparse Expert Models.

5. AI Application Engineer — the most mainstream LLM role since 2023

AI Application Engineers (often called LLM Application Engineers) integrate general-purpose LLMs into business systems: prompt engineering, RAG, Agent workflows, fine-tuning (LoRA), evaluation set construction, cost and latency optimization. This role doesn't require you to train LLMs from scratch, but it does require you to understand model capabilities and know how to compose them with engineering. These are the most numerous roles with relatively friendly entry barriers, making them the primary battlefield for most career-switchers.

• Core skills: Python, LLM APIs and open-source model deployment, RAG in practice, evaluation practice, Agent frameworks, and product-level engineering.
• Typical salary: ~300K–700K RMB/year in tier-1 Chinese cities (experienced and scarce talent command significant premiums); overseas ~$180K–300K/year.
• Related pages on this site: RAG principles in RAG: Retrieval-Augmented Generation, prompt engineering in Prompt Engineering, and hands-on practice in RAG in Practice.

6. Agent Engineer — teaching models to "work on their own"

Agent engineers' core mission is to transform LLMs from "question-answerers" into "doers": designing tool calling (function calling), planning loops (ReAct), memory mechanisms, multi-agent collaboration, and connecting to real-world APIs. The 2025 market has split noticeably: some companies use Agent frameworks for complex workflows, while others retreat to "single-turn deterministic pipelines." An Agent engineer's real value lies in knowing when to let the model act autonomously and when to hardcode logic.

• Core skills: LLM-based Agent architectures, prompt engineering, tool calling and protocols (e.g., MCP), safety & risks (prompt injection), and systems engineering.
• Typical salary: ~350K–800K RMB/year in tier-1 Chinese cities; overseas ~$180K–320K/year.
• Related pages on this site: Agent case studies and boundaries in LLM-based Agents; risk mitigation in Safety & Risks.

7. Evaluation Engineer — the scarce "gatekeeper"

Evaluation engineers score every change to a model: building benchmark and custom sets, running offline batch evaluations, designing LLM-as-a-judge, performing regression tests to prevent degradation, and monitoring post-deployment metrics and user feedback. This is a widely recognized scarce role in LLM teams — most teams over-invest in training and under-invest in evaluation, until they hit the point where "we've made ten iterations and nobody knows if it actually got better." Evaluation work demands rigorous statistical thinking and vigilance against data contamination.

• Core skills: Evaluation & benchmarks systems, benchmark profiles (MMLU, GSM8K, HumanEval, etc.), hallucination and safety testing, evaluation automation (lm-eval-harness / OpenCompass), and data analysis.
• Typical salary: ~300K–700K RMB/year in tier-1 Chinese cities; overseas ~$180K–280K/year.
• Related pages on this site: Benchmark and dataset profiles in Datasets & Benchmarks; hands-on practice in Evaluation Practice.

8. Data Engineer · LLM Direction — data quality sets the model's ceiling

"Data is the model" — the quality of pretraining corpora directly determines the base model's ceiling, and post-training data (instructions, preferences) determines the end-user experience. Data engineers handle corpus collection, cleaning, deduplication (MinHash), filtering, mixing, compliance review, and managing labeling teams. Since 2024, synthetic data has become a hot area: using models to generate high-quality training data.

• Core skills: Full pretraining data pipeline, tokenization and vocabulary, big data tools (Spark/Flink), data pipeline engineering, and copyright & compliance awareness.
• Typical salary: ~250K–600K RMB/year in tier-1 Chinese cities; overseas ~$150K–250K/year.
• Related pages on this site: Data and benchmark profiles in Datasets & Benchmarks.

4. Skill Radar: Algorithms · Engineering · Applications · Math

Plot the seven roles on a four-dimensional coordinate system, and their personalities become immediately apparent. The four dimensions are:

                Algorithms (Transformer internals / training details / tuning)
                          ▲
                         /|\
                         │
    Math ───────────────┼──────────── Engineering
    (linear algebra /    │           (code quality / deployment /
     probability /       │            MLOps / systems)
     derivations / stats)│
                         │
                        \|/
                Applications (RAG / Agent / product / business metrics)

Each dimension is scored 1–5 (1 = basically not needed, 5 = core capability for the role):

Role	Algorithms	Engineering	Applications	Math	Role Personality
Algorithm Engineer (LLM)	4	3	3	3	Full-stack generalist
NLP Algorithm Engineer	4	3	2	3	Model-builder
LLM Training Engineer	5	4	1	3	Scientist + engineer
Inference Optimization Engineer	3	5	1	3	Systems engineer
AI Application Engineer	2	4	5	2	Engineer + product
Agent Engineer	2	4	5	2	Engineer + product
Evaluation Engineer	3	3	3	4	Analytical / rigorous
Data Engineer (LLM)	2	4	2	2	Pipeline engineer

How to use this table for choosing

Don't just look for "which row has the highest number." Look at which column you can see yourself deep-diving long-term: someone strong in algorithms tolerates the boredom of loss curve diagnostics, someone strong in engineering enjoys the satisfaction of a stable system, someone strong in applications gets fulfillment from "users are actually using this," and someone strong in math enjoys the certainty of derivations. Put "which type of work would I prefer to do long-term" above "which role is currently most popular."

5. Market Trend Analysis (2023–2025)

Phase	Market Signal	Impact on Roles
2022.11–2023	ChatGPT went viral; LLM application roles went from zero to many	Prompt engineering and RAG roles exploded; "just know how to call the API" was enough to get in
2023–2024	Open-source models (Llama/Qwen/DeepSeek) caught up to closed-source; surging demand for private deployment	Rising demand for deployment, fine-tuning, and inference optimization engineers; "only know how to call APIs" starts depreciating
2024–2025	Agent concept matures and diverges; training costs remain high, training roles consolidate at top companies	Application layer continues expanding; evaluation and safety roles stay scarce; engineering (deployment, evaluation, observability) gains weight
2025 outlook	Inference costs keep dropping; long context and multimodal become standard	Multi-disciplinary talent who "understands evaluation, costs, and systems" becomes more valuable

Key facts to factor into your decisions (as of dataAsOf: 2025-08):

• Pure "hyperparameter-tuning" general algorithm roles are declining, while "engineering + LLM application" hybrid demand is rising; the proportion of JDs mentioning LLM keywords is climbing fast.
• Barriers for training roles keep rising. Open-source ecosystems have reduced the need to "train from scratch," so most companies are shifting to fine-tuning and post-training.
• "Invisible roles" like evaluation, safety, and data quality are gaining recognition — they're insulated from single-wave technology hype.
• Vertical domains like NLP / recommendation / search haven't disappeared, but all of them are layering LLM capabilities on top.

Don't take the "role title" at face value

The same title can mean completely different things at different companies: at Company A, an "Algorithm Engineer" might spend all day tweaking prompts, while at Company B, an "AI Application Engineer" might spend all day writing data processing code. Reading actual JDs is always more important than reading role titles — which is exactly why the first stop in this module is the JD List, not a "role encyclopedia."

6. Decision Framework: Which Direction Should You Choose?

There's no standard answer, but there are standard questions. Go through the three checklists below, and your answers will converge to one or two roles.

1. By background: Where are you standing now?

• [ ] CS/Software background, years of coding, want to get into LLM → prioritize AI Application Engineer / Agent Engineer / Inference Optimization Engineer
• [ ] Math/Statistics/Physics background, comfortable with derivations but limited engineering → prioritize Algorithm Engineer / Evaluation Engineer
• [ ] Classical NLP / Search / Recommendation background → prioritize NLP Algorithm Engineer, layering LLM capabilities onto your existing domain
• [ ] Fresh graduate with no industry experience, want the fastest path to interviews → AI Application Engineer has the most open JDs in China and is the most universal entry point
• [ ] Already have data/big data engineering experience → prioritize Data Engineer (LLM direction) — engineering experience is a clear advantage

2. By interest: What kind of "completion feeling" do you enjoy?

• [ ] "The training loss finally went down" → Training & algorithms: training engineer / algorithm engineer
• [ ] "This service is 99.9% available, and latency dropped by half" → Systems: inference optimization engineer
• [ ] "I assembled an LLM into a feature that real users are using" → Product + tech: AI application engineer / Agent engineer
• [ ] "I designed an evaluation system that stopped the team from guessing" → Analytical: evaluation engineer
• [ ] "I built the entire corpus pipeline from crawling to production" → Data: data engineer

3. By trend: How much certainty are you willing to trade for?

• Want stability → evaluation / data / application engineering, less affected by single-wave hype
• Want speed → Agent / inference optimization, concentrated hype, fast raises but volatile
• Want long-term horizon → training & algorithms, high ceiling but few openings and high barriers

Look at the three sets of results together: if all three point to the same direction, go for it. If they conflict, use background as your safety net, interest as your compass, trend as your weight — background determines whether you can get in the door, interest determines how far you can go, and trend only determines how crowded the doorway is.

7. Content Map: What Each of the Five Pages Does

The career module has five pages, each corresponding to a step in the job-hunting process:

Page	One-Liner	When to Use
JD List	JD templates and keyword radar organized by role type, breaking down high-frequency requirements and thresholds	Day 1 of job-hunting: first learn what the market wants
Knowledge Breakdown	Maps JD high-frequency requirements to knowledge points and site pages, showing "how will you be asked in an interview"	After understanding the market: find your gaps
Resume Analysis	Resume review standards from the interviewer's perspective, with rewrite examples for fine-tuning/RAG/Agent/evaluation/deployment projects	When preparing to apply: turn experience into evidence
Interview Questions	30+ high-frequency questions organized by topic, with key reference answers and a self-test checklist	One week before interviews: final self-assessment

The reading order is the main thread, not an option list

The Interview Prep Path has already arranged the five pages as a battle plan: JD List → Knowledge Breakdown → Resume Analysis → Interview Questions. The output of each page feeds into the next. Jumping around will cost you most of the value.

8. Two Reminders

1. Check the dataAsOf date before trusting numbers

The roles, salaries, JD requirements, and technical keywords in this module all have expiration dates. The LLM field has an extremely short half-life — a role's JD can be rewritten twice in a year, and salary ranges fluctuate with market conditions. That's why every time-sensitive page in this module has a dataAsOf field (the cutoff month for data). Before quoting specific numbers, check that date:

• Within one quarter of the current date → treat as current reference
• More than one quarter old → treat as "trend reference," and verify with official sources
• Conceptual content (attention formulas, KV cache) decays very slowly and is not subject to this constraint

2. JD descriptions vs. actual work: two versions of the world

A JD describes "what they wish you were"; actual work is "what they actually need you to be." There's a systematic gap between the two:

• JD says "proficient in deep learning," reality is probably lots of data cleaning and evaluation scripts — model tuning is only a small part of the job
• JD says "optimize model performance," reality is business metrics and cost — improving accuracy by 2 points but doubling latency won't pass review
• JD says "independently responsible for the LLM system," reality is you're one piece of a massive system — describing "the part you owned" clearly is far more credible than bragging about "the whole system"
• JD says "Agent experience preferred," reality is the team hasn't even figured out what they want yet — honestly discussing your understanding of this uncertainty is itself a plus

There's only one response: do intelligence gathering on your target team before the interview (read interview experiences, talk to current employees, try the product), treat JDs as hypotheses, and verify with research. The Interview Questions page has a category of "reverse questions for the interviewer" that exists specifically for this.

9. Further Reading

Continue within the site

• Interview Prep Path — the complete battle plan that the career module is part of
• System Architecture Anatomy — a page map of the full LLM system lifecycle, so you can see where each role fits in the chain
• Knowledge Breakdown — the bridge from roles to knowledge points; generates your personal study plan
• Glossary — come back to for consistent definitions when reading JDs and interview questions

References

• levels.fyi — compensation database for US tech companies, searchable by company/level/city; reliable source for overseas salary data
• Stack Overflow Developer Survey — annual developer survey with AI/LLM tool adoption rates and salary distribution
• BOSS Zhipin — China's leading job platform; check real-time LLM role JDs and salaries
• Zhaopin — domestic recruitment platform with industry heat reports and demand data
• OpenAI Careers — official career page for a top LLM company; see real overseas JDs