Opinion pieces about "what skills AI engineers need" are everywhere. This one is different: it's based on 808 active AI engineering job listings on SuperAIDevs, analysed this week.
Here's what employers are actually hiring for right now — not what bloggers think sounds important.
The Data
We counted how many job listings mention each technology across our active listings. Here are the top 15:
| Skill | Jobs mentioning it | % of listings | |-------|-------------------|---------------| | LLMs | 643 | 80% | | Python | 584 | 72% | | RAG | 437 | 54% | | LangChain | 334 | 41% | | AWS | 324 | 40% | | Azure | 296 | 37% | | OpenAI API | 244 | 30% | | LangGraph | 224 | 28% | | GCP | 197 | 24% | | Docker | 185 | 23% | | Kubernetes | 178 | 22% | | Anthropic | 172 | 21% | | PyTorch | 134 | 17% | | TypeScript | 136 | 17% | | SQL | 132 | 16% |
A few things jump out immediately.
What This Actually Means
LLMs + Python is the non-negotiable baseline
80% of jobs mention LLMs, 72% mention Python. If you don't have both, you're filtered out before a human reads your CV. This isn't surprising — but the gap between these two and everything else tells you where the floor is.
"LLMs" in job listings means: understanding how language models work, how to prompt them effectively, how to evaluate their outputs, and how to integrate them into production systems. It's not just "I've used ChatGPT."
RAG is now a core skill, not a specialisation
54% of listings mention RAG (Retrieval-Augmented Generation). A year ago this was advanced. Now it's table stakes. If you haven't built a RAG pipeline from scratch — chunking, embedding, vector search, retrieval, re-ranking, generation — put it on your list for this month.
The canonical stack: LangChain or LlamaIndex for orchestration, Pinecone or pgvector for storage, OpenAI or Anthropic for generation.
LangChain vs LangGraph
LangChain appears in 41% of listings, LangGraph in 28%. These aren't competing — LangGraph is built on top of LangChain for stateful, agentic workflows. Employers want both. If you know LangChain, learning LangGraph is a weekend project. If you don't know either, start with LangChain.
Cloud is mandatory, not optional
AWS (40%), Azure (37%), GCP (24%). Combined, some cloud platform appears in the majority of senior listings. The era of "I'll learn cloud later" is over for AI engineers — you need to be able to deploy what you build.
AWS is the safest single bet (most common), but if you're interviewing primarily at enterprises, Azure is likely more relevant (Microsoft's OpenAI partnership puts Azure in most enterprise AI stacks).
The surprise: Go is in 21% of listings
169 jobs mention Go. This reflects the reality of production AI infrastructure — Python is for models and pipelines, Go is for the services that wrap them when you need performance and low latency. Not required for most roles, but a genuine differentiator if you have it.
The Three Tiers
Based on the data, here's how to think about prioritisation:
Tier 1 — Required for most roles (learn these first if you don't have them)
- Python (proficient, not just scripting)
- LLM fundamentals (prompt engineering, evaluation, fine-tuning basics)
- RAG pipelines (build one end-to-end)
- At least one major cloud platform
Tier 2 — Required for senior roles, expected in 12 months
- LangChain + LangGraph (agent orchestration)
- OpenAI API + Anthropic API (both, not just one)
- Docker (containerise your work)
- Vector databases (Pinecone, pgvector, Weaviate)
Tier 3 — Differentiators
- Kubernetes (for MLOps/platform roles)
- PyTorch (for roles closer to model development)
- TypeScript (for full-stack AI engineers)
- Go (for performance-critical infrastructure)
The Skills Nobody Lists but Everyone Wants
Three things that don't show up cleanly in job postings but come up constantly in interviews:
Evaluation and evals frameworks. How do you know if your RAG pipeline is actually good? Can you set up automated evaluation using LLM-as-judge? Can you measure hallucination rates? This is where most candidates fall down.
Cost and latency awareness. GPT-4o at 100k tokens/day costs real money. Senior engineers think about caching strategies, prompt compression, model selection for different tasks (don't use GPT-4o for a classification task that GPT-4o-mini handles fine at 10x less cost).
System design for AI. How do you handle a prompt injection attack? What happens when the LLM returns malformed JSON and your downstream parser breaks? How do you handle rate limits at scale? Production AI engineering is full of failure modes that don't exist in classic software.
What to Build This Month
The fastest path to being hireable for AI roles isn't another tutorial. It's a project you can point to.
Build one of these:
- A RAG system over a real document corpus (your own PDFs, a public dataset)
- An agent that uses tools (web search, code execution, API calls) to complete a multi-step task
- An automated evals pipeline that measures your model's performance over time
Put it on GitHub with a good README. Write a short post about what you learned. That combination — working code + written reflection — stands out in a field full of tutorial followers.
When you're ready to apply, browse 800+ AI engineering roles on SuperAIDevs — every listing is tagged with the specific stack so you can filter to roles that match what you've built.