Reality check · GenAI experience
What Companies Mean When They Say "Generative AI Experience"
Here is a line that now appears in many Indian job postings for developers, analysts and even product managers: "Hands-on experience with Generative AI (LLMs, RAG, prompt engineering) is a plus." If you have used ChatGPT every day for a year, does that count? Usually not. If you have never held a job with "AI" in the title, are you ruled out? Also no. The requirement is about whether you have built something with generative models and dealt with the problems that come with it. This page reads that line piece by piece and shows how to earn it.
Reading the requirement word by word
"Hands-on"
The hiring manager wants evidence that you have written code or configured a system that calls a model, not just read about it. Your GitHub repository, a deployed demo or a clear project write-up is the evidence.
"LLMs"
You should know how to call a large language model through an API (from OpenAI, Anthropic, Google, or an open model served locally), control output with parameters such as temperature, and handle practical limits like context length, rate limits and cost per request.
"RAG"
Retrieval-augmented generation means fetching relevant passages from your own documents and giving them to the model so it answers from facts. This is the most common GenAI pattern in companies, because most business questions are about the company's own data: policies, product manuals, contracts, tickets.
"Prompt engineering"
Here it means writing structured, testable instructions and output formats inside an application, not clever one-off chat prompts.
"Is a plus" versus "required"
When it says "plus", a small but real project is often enough to stand out. When it says "required" in an AI engineer posting, expect interviews to go deep into how you evaluated and deployed your system.
The levels of GenAI experience
| Level | What you have done | Counts for |
|---|---|---|
| 1. Tool user | Used ChatGPT, Copilot or Gemini to speed up your own work | Nice to mention for non-tech roles; not "experience" for technical roles |
| 2. API builder | Called a model API from code, handled errors, parsed structured output | "Plus" requirements in developer and analyst roles |
| 3. RAG builder | Built retrieval over documents with embeddings and a vector store, and checked answer quality | Most GenAI developer roles |
| 4. Agent / tool-use builder | Let a model call functions or APIs to complete multi-step tasks, with limits and logging | AI engineer and automation roles |
| 5. Production owner | Evaluated, deployed and monitored a GenAI feature with real users, cost and safety controls | Senior and lead AI roles |
Fine-tuning models sits alongside levels 4 and 5. It is valuable but less common than people think; most companies get further with good retrieval and prompts before they fine-tune anything.
What usually does not count
- A certificate from a short GenAI course with no project attached.
- A chatbot demo that is a thin wrapper over a public chat interface, with no data, testing or design decisions of your own.
- A list of tools on the resume ("LangChain, LlamaIndex, Pinecone") that you cannot explain in an interview.
- Copying a tutorial repository without changing the problem, the data or the approach.
Interviewers spot these quickly by asking one follow-up: "What went wrong, and how did you fix it?"
What retrieval looks like in code
The core of RAG is small enough to understand in one sitting. This snippet finds the passage most relevant to a question using an open embedding model:
from sentence_transformers import SentenceTransformer, util
docs = [
"Returns are accepted within 30 days of delivery with the invoice.",
"We deliver to most pin codes in India within 5 working days.",
"Refunds are credited to the original payment method.",
]
model = SentenceTransformer("all-MiniLM-L6-v2")
doc_vecs = model.encode(docs, convert_to_tensor=True)
question = "How many days do I have to send a product back?"
q_vec = model.encode(question, convert_to_tensor=True)
scores = util.cos_sim(q_vec, doc_vecs)[0]
best = int(scores.argmax())
print(docs[best])
It prints the returns sentence, even though the question uses different words ("send back" instead of "returns"); that is what embeddings add over keyword search. In a full system, the best passages go into the model's prompt along with an instruction to answer only from them. The hard part, and the part interviewers ask about, is everything around this: splitting long documents into chunks, handling questions with no answer in the documents, and measuring whether answers are correct. RAG, agents, APIs and vector databases explained covers these pieces in more depth.
Three projects that create real experience
- Document assistant with evaluation. Pick a set of public documents you understand, such as your university's ordinances or a government scheme's FAQs. Build RAG over them, write 30 test questions with expected answers, and report how many it gets right before and after your improvements. The RAG customer support assistant project is a good template.
- Structured extraction. Use an LLM to turn messy text (invoices, job postings, resumes with fake data) into clean JSON fields, validate them in code, and measure the error rate on a sample you labelled yourself.
- A small agent with guardrails. Let a model call two or three functions, for example look up an order, check stock, draft a reply, with a hard limit on steps and a log of every call. Show what happens when the user asks for something outside scope.
If you already have a job, the best experience is a small GenAI improvement to your own team's work, done with your manager's approval and without sending confidential data to outside services.
How to write it on your resume
Weak bullet:
Worked on Generative AI using ChatGPT, LangChain and vector databases.
Stronger bullet (for a hypothetical project):
Built a RAG assistant over 120 pages of exam regulations (Python, open embeddings,
FAISS); created a 30-question test set and improved correct answers from 18 to 26
by changing chunk size and adding a "not found" rule.
The second version names the problem, the stack, a measurement you can defend and a decision you made. Only write numbers you actually measured. Then be ready to explain each part in an interview; preparing for AI engineer interviews lists the questions that typically follow.
Where to go from here
If the terms in this page are new, start with what you need to learn about LLMs, then build the first project. If you are wondering whether using ChatGPT well is enough on its own, ChatGPT skills vs AI engineering skills explains the gap.
Frequently asked questions
Does using ChatGPT at work count as Generative AI experience?
For non-technical roles it can be worth mentioning with a concrete result. For developer, analyst or AI roles, companies usually mean building with model APIs, retrieval or agents, not using a chat tool.
Can a fresher have Generative AI experience?
Yes. Well-documented projects that call models, use your own data and measure results count as hands-on experience, especially when the posting lists GenAI as a plus.
Do I need to fine-tune models?
Not for most roles. Retrieval, prompt design and evaluation cover most company use cases. Fine-tuning is useful to understand and matters more for specialised ML roles.