You accepted a role titled "GenAI Engineer" or "AI/ML Engineer". Nine months in, your actual work is hand-labelling text, drawing boxes on images, checking model outputs against a rubric, or filling status decks. You have written almost no code. Your resume now says AI Engineer and your GitHub shows nothing from the job, and interviewers at the next company find the gap in four minutes.
First job, taken under time pressure at the end of a placement season, often with a service agreement attached. The job description mentioned Python, LLMs and RAG, and it also mentioned "support annotation and quality review of training data" in a lower bullet nobody read carefully. No question was asked about the daily work before the offer was signed. Linked module: 6.1 (How Indian companies actually hire).
Titles in Indian job postings inflate for two ordinary reasons: a good title attracts more applicants at the same pay, and internal grade names rarely match what the market calls the work. Neither is fraud, and neither is your friend.
The evidence was in the posting all along, in the task verbs. Verbs like design, build, train, deploy, evaluate, debug describe engineering. Verbs like support, review, tag, annotate, verify, escalate, coordinate, prepare describe operations work around a model somebody else built. A posting whose engineering verbs sit in the "nice to have" section and whose operations verbs sit in the "responsibilities" section is an operations job with an engineering title.
Two more signals were available before signing and were never checked. First, the team: how many engineers, who reviews code, what has this team shipped. Second, the stack: which model, which framework, which repository you would commit to in month one. A hiring manager staffing an engineering role answers both instantly. A vague answer to both is the diagnosis.
Note what is not a diagnosis. Data work is real, necessary and sometimes a fine entry point; a labelling-heavy first year at a firm that promotes internally can beat unemployment. The failure here is not the work — it is accepting a role whose real content you never asked about, and then discovering it after a bond made leaving expensive.
Before you sign anything, ask five questions and write the answers down.
Then read the contract terms with the same care: service agreement length, penalty amount, notice period, and whether the role or location can be changed at the company's discretion. If you are already inside such a role, do not quit in anger. Extract every transferable asset the job does offer — data quality judgement, evaluation rubrics, domain knowledge, the vocabulary of a real production system — and build the engineering evidence outside work, in public, so the next application carries proof the job did not give you.
Read every job description bottom-up: responsibilities first, title last. Mark each line SIGNAL or NOISE, and summarise the role from the task verbs alone before you look at the title again. Ask the five questions in every process, not only the ones you are unsure about — asking them is itself a signal that reads well. Keep your own definition of the role you want written down, so a flattering title cannot quietly replace it.
Q: "What does an AI engineer do here, in your view?" This is your opening to ask the five questions without sounding suspicious. Answer briefly with the task verbs you read in their own posting, then ask what a typical week looks like and what you would commit code to first. Candidates who do this are remembered as serious, and the answers you get decide whether you should accept — which is the half of the interview most freshers forget they are also conducting.