# AI Engineering Full Roadmap vs 100-Day Sprint: In-Demand Skills, Job Market Data & Curriculum

[Watch on tuber](https://tuber.guzus.xyz/video/dQgsHROf8Cw/summary) · [YouTube](https://youtu.be/dQgsHROf8Cw)

**Score: 64/100 · skim**

The video presents CampusX's AI engineering roadmap. The instructor argues that AI engineering, meaning building, evaluating, securing and deploying applications on top of existing LLMs rather than training models, is the fastest-growing tech job. He walks through a full 16-level roadmap (about 286 two-hour sessions, 8-12 months) and then a condensed "100 Days of AI Engineering" version that he says keeps roughly 70-90% of the value. RAG, agents and evaluation are the areas he treats as essential.

- **0:03-2:11: Session goals.** He promises a detailed roadmap built from about two years of teaching in this field, starting with LangChain in January. He also wants to explain CampusX's course plan for the next 6-12 months. He mentions frustration with repeated comment requests for topics like fine-tuning and says the company follows a planned vision, not random course releases.

- **2:11-6:00: What an AI engineer is.** He separates this role from data scientists, data analysts and the people who train foundation models at labs like OpenAI and Anthropic. An AI engineer takes an existing LLM and builds applications on it: RAG systems, agents and multimodal apps. They also deploy, secure, evaluate and manage those apps. He warns that the title isn't standardized, so many job posts also ask for classical ML, deep learning or fine-tuning (for example, an XGBoost model on tabular data). He expects the role to become more clearly defined over time.

- **6:00-10:17: Why the role is growing.** Few companies need to build their own models, because that is expensive and requires rare skills. Almost every business, though, can plug LLM "intelligence" into its existing problems. His examples:
  - Customer support teams of around 200 people, with constant training and attrition costs, could be replaced by voice AI agents.
  - Vision models could power driverless cars at companies like Uber.

  He cites Google Trends: the term was flat until ChatGPT in 2022, then rose sharply, peaked and dipped slightly, but the trend is still upward.

- **10:17-16:17: The India AI Workforce Report.** He presents a roughly 44-page report from a listed staffing company (on the NSE) that he says has placed 9-10 lakh people, though the company name doesn't come through clearly in the transcript. The report's main numbers:
  - About 920K AI professionals in India.
  - About 3.5 lakh AI-skill job postings in 90 days, which he says could mean over 3 lakh jobs a year if the pace holds.
  - AI hiring is spreading beyond ML into software engineering, cloud and cybersecurity.
  - Demand is concentrated at 3-5 years of experience, with a slow market for freshers.
  - The fastest-growing roles include agentic AI developer, agentic AI/MCP engineer, RAG and agentic AI lead, and GenAI solutions manager, with about 260% year-over-year growth.
  - Forward-deployed engineer (FDE) roles are emerging, with an estimated 12-15K more needed.

  He promises a full discussion of the report the next day.

- **16:17-21:25: Layoffs and mindset.** He acknowledges fears from events like Oracle firing 30,000 people. He agrees that traditional roles, especially support delivery and entry-level coding, are highly exposed to automation. The same report projects 2.3 million or more new roles by 2030 in areas like AI evaluation, AI security and AI governance (operating AI under different countries' rules). He says these tasks can't simply be handed to AI. Teams will be small, so the goal should be to become strong enough to land one job. He argues that people who follow AI closely have an advantage in foresight over people in fields like sales who don't.

- **21:25-28:00: How the roadmap is structured.** There are two versions:
  - **Comprehensive roadmap:** a "superset" of 16 levels (0-15), about 286 sessions, 8-12 months.
  - **Accelerated roadmap:** the 100-day version, which he says is enough to crack interviews in 4-6 months.

  Resources are mostly CampusX courses, because he only recommends material he has tested and few trusted external resources exist yet. The roadmap is hosted on Vercel and linked from the CampusX site, with FDE and data science roadmaps planned. Each level lists prerequisites, modules, projects and related courses, plus two 0-10 scores: interview weight and on-the-job weight.

- **28:00-34:35: Level 0, prerequisites.** These are software engineering fundamentals, not heavy math, because you aren't training models. The list:
  - Python: syntax, data structures, OOP, Pydantic data models, file handling, environment variables, async (needed for RAG and agent calls), modules, and testing and debugging for CI.
  - APIs: web, HTTP, REST and JSON, plus FastAPI (or Flask, not both).
  - SQL and database schema design, for example to store chatbot memory.
  - Git and GitHub.
  - Structured, spec-driven use of AI coding tools.
  - Docker basics.

  He says some companies screen candidates on API and backend questions. His advice is to spend at most one month here and not chase perfection.

- **34:35-42:36: Level 1, LLM fundamentals (interview weight 10).** He added this level later because most engineers jump straight into building. It covers:
  - Next-token prediction and autoregressive generation.
  - Tokenization (all costs are measured in tokens).
  - Transformer basics: embeddings, positional encoding, attention.
  - Pre-training and post-training/alignment at a conceptual level.
  - How inference works: KV cache and where costs come from.
  - Sampling settings like temperature, and why the same prompt gives different outputs.
  - Reasoning models and when their cost and latency are worth it.
  - Context windows, the model landscape (proprietary vs open source), benchmarks, and how to choose a model.

  His example is prompt caching: an engineer who understands it saves money. Interviewers often test tokens, temperature, hallucinations, KV cache and context limits. No CampusX course exists yet; he plans to teach one within 1-2 months, after his current Google ADK course.

- **42:36-49:48: Levels 2-3, LLM APIs and prompt engineering.** Level 2 (interview 8, job 10, 19 modules) is mandatory. It covers:
  - API calls via OpenAI or LangChain, and managing conversation state.
  - Streaming, structured output, and function and tool calling.
  - Multi-step workflows, parallel calls and batching, retries and fallbacks when a provider's server fails.
  - Observability for cost and latency.
  - LangChain is recommended; LlamaIndex is optional.

  The courses are GenAI with LangChain, Ollama for open-source models, and a Gemini course, which he doesn't push because the material is tied to one ecosystem. Level 3 is prompt engineering, which he defines as writing system prompts for software, not consumer prompt tips. It's rarely asked in interviews but used constantly at work. Topics include zero- and few-shot prompting, prompting reasoning models, system prompt design, output control, reliability and debugging, prompt versioning (so you can roll back after 10-20 edits) and prompt-injection defense.

- **49:48-1:00:52: Levels 4-5, RAG and AI agents (the core).**
  - **RAG (interview 10, job 10, about 27 lectures).** He calls this the most important module and denies that RAG is dead, because most industry projects are still RAG over internal documents, and people who cracked interviews say they were grilled most on it. Basics: ingestion, chunking, embeddings, vector databases, semantic search, and sparse/hybrid retrieval (BM25). Advanced: reranking, query transformation, citations, Self-RAG, CRAG, adaptive RAG, and frameworks. Strong RAG skills can make interviewers forgive weaker areas.
  - **Agents (interview 9, job weight lower, about 22 lectures).** Most companies actually build fixed workflows. His example is an email classified by an LLM: that is a workflow, not an agent, because a true agent decides its own steps. Topics: agent = LLM + tools + memory; ReAct and planning agents; tools and MCP; memory; multi-agent architectures and A2A; LangGraph; sandboxed execution (his file agent was confined to a test folder, and startups now sell sandboxes); and failure recovery for long-running agents.
  - **Libraries:** he recommends LangGraph for control and Google ADK for industry adoption over CrewAI, Agno or AutoGen. An agent memory course is coming in 2-3 months. DSA matters somewhat in interviews depending on the company but almost never on the job.

- **1:00:52-1:12:06: Levels 6-8, context engineering, evaluation and security.**
  - **Context engineering (interview about 8, higher for senior roles).** This means managing what goes into a limited context window in real time. His example is a coding agent that loads many files and degrades once it breaches the limit. Topics: summarization and compaction, moving state into memory, "lost in the middle" (models remember the start and end but forget the middle), growth from tool outputs, context caching, and harness engineering. He explains a harness as the custom code that glues an LLM to tools and data, for example an eval harness built around an LLM judge, and cites OpenClaw as an example.
  - **LLM evals (job weight 10).** He compares this to software testing splitting off into QA and predicts dedicated LLM application testers. Teams found that deployed RAG chatbots didn't perform as expected. At least one interview question is almost always "how do you know your RAG or agent works?" The level covers LLM, RAG and agent evals, and a four-month CampusX course already exists.
  - **AI security (interview weight lower for now).** Topics: threat modeling (diagramming components and their attack surfaces), prompt injection, data exfiltration, jailbreaks, agent-specific attacks, PII leakage, red teaming and governance. A six-module guardrails course is due by October 31, which he sees as an edge over other candidates.

- **1:12:06-1:19:08: Levels 9-10, AI system design and LLMOps.**
  - **AI system design (interview 8, job 8).** It has its own interview round and is critical for senior roles. Topics include model routing and gateways (cheap models for simple tasks, stronger ones for complex tasks), inference optimization, multi-level caching and guardrails. There's no CampusX course yet (planned within about 4 months). He recommends Chip Huyen's *AI Engineering* book, the only external resource in the roadmap.
  - **Airport analogy:** design is planning the runway, terminals, passenger load and security before building; LLMOps is running daily operations, meaning landing flights and screening passengers.
  - **LLMOps (interview 7, job 9).** Covered in Himanshu's course over roughly the last 3-4 months: prompt management, evaluation, guardrails, Docker, CI/CD and post-deployment monitoring.

- **1:19:08-1:27:13: Optional Levels 11-15.**
  - **Managed AI platforms** (Bedrock, Azure Foundry, Vertex AI): only matter if your company uses that cloud, likely covered in December after a poll.
  - **Multimodal AI** (vision, audio, video, real-time and edge): low priority unless your domain needs it.
  - **Fine-tuning:** about 7 for interviews, but rarely needed at work, since good prompting and RAG usually suffice. Topics are LoRA-style techniques, alignment, distillation, and evaluating and serving tuned models. He plans to release LLM 101, AI system design and fine-tuning courses within about six months, by roughly February.
  - **Projects** (interview 10): you get grilled on them; 5-6 suggested projects will be built out from December.
  - **Interview prep:** sections by topic plus behavioral, backed by an existing 6-7 session interview course.

  He stresses that nobody can realistically finish all of this; the full map exists to show the boundaries of the field.

- **1:27:13-1:36:42: The "100 Days of AI Engineering" accelerated roadmap.** It drops to about 8 levels and works through one evolving project:
  - Level 0 is removed.
  - LLM fundamentals are trimmed (transformer internals dropped).
  - APIs are trimmed (direct provider calls, retry logic and LlamaIndex removed).
  - Prompting is reduced to essentials.
  - RAG is learned by building one app, making it advanced, then evaluating and securing it in the same module.
  - That app is then turned into an agentic RAG app with MCP, memory and multiple agents, again with evaluation and security built in.

  Folding evals and security into the build removes two separate modules. A "Shipping to Production" level (about 15 days) combines system design and LLMOps: front end, packaging, deployment, monitoring, scaling, and trying an open model or fine-tuning. It ends with 2-3 projects using the same flow and the 100 most-asked interview questions. A live or YouTube cohort may follow, but nothing is promised.

- **1:36:42-1:39:49: Closing.** The roadmap is CampusX's "North Star": Himanshu, Roopendra, Ajay and new hires are each assigned courses, projects or interview content, aiming to cover the full map within about six months. He says he isn't trying to sell and invites people to submit good free resources, which CampusX will test and add to a new section. The final minutes preview the next day's session on the report and career questions.

**Practical takeaways**
1. Put RAG, agents and LLM evals first. They carry the most interview weight and reflect what companies actually build; be ready to explain how you'd evaluate any app you made.
2. Learn LLM internals (tokens, KV cache, temperature, context limits, prompt caching) even though you won't train models. Interviewers use them to judge depth, and they drive cost and latency decisions on the job.
3. Spend no more than a month on prerequisites (Python incl. async, FastAPI, SQL, Git, Docker), and treat fine-tuning, multimodal and managed cloud platforms as optional unless your target company needs them.
4. If time is short, follow the 100-day path: build one RAG app into an agentic system, evaluating and securing it as you go, then ship it to production. That project becomes your portfolio piece.
5. Read Chip Huyen's *AI Engineering* for system design, which matters more as you move toward senior roles.

## Audience signal

- No high-signal audience themes available.