Resume for AI and GenAI Jobs (Without Faking Machine Learning)

The loudest job titles on LinkedIn right now are variations of AI engineer, GenAI developer, and “prompt engineer.” Half of the postings are real product work. The other half are old backend jobs with ChatGPT sprinkled into the description. Your resume has to survive both: a recruiter who searches for Python and LangChain, and an engineer who will ask you to explain a retrieval pipeline on a whiteboard. This guide is for people who have actually shipped something with models — even a messy internal tool — and for people who have only used ChatGPT as a writing assistant and should not pretend otherwise.

I read a lot of first drafts that open with “passionate about artificial intelligence and transforming the future of work.” That sentence could sit on a thousand PDFs. It tells nobody whether you fine-tuned anything, wrapped an API, evaluated answers, or just pasted prompts into a chatbot. In 2026 the market is picky in a boring way. Companies want people who can take an existing model, constrain it with data the company already has, and keep it from lying to customers. That is product engineering with extra failure modes. Write the page that way.

What these roles actually hire for

Strip the branding and most GenAI job posts ask for four clusters. First, software: Python or TypeScript, APIs, git, tests, some cloud. Second, data plumbing: chunking documents, embeddings, vector stores, eval sets, logging. Third, product sense: when not to use a model, how to cite sources, how to fail closed. Fourth, cost and safety: token budgets, PII, prompt injection, rate limits. Research scientist posts exist, but they usually want papers, CUDA, and a different CV. If you are applying through MineResume, you are almost certainly in the first bucket — applied AI, not a lab.

India’s service companies and product startups use slightly different language. Service firms still write “GenAI developer – LangChain, OpenAI, Azure OpenAI.” Startups write “build RAG on our support corpus.” US and EU posts add evaluation harnesses and “you will be on-call when the bot hallucinates in production.” Translate your work into their nouns, but only if you did the work. Calling yourself an ML engineer because you completed a Coursera neural-net week will collapse in round one when they ask about gradients. Call yourself a software engineer who shipped an LLM feature. That title is hireable and defensible.

Projects that count versus tutorials that do not

A notebook that classifies flowers is a course artifact. A repo that takes a folder of PDFs, splits them, stores embeddings, answers questions with citations, and has a README with setup and a known failure case is a project. Recruiters cannot run your notebook in the ATS. They can open a GitHub link from the header and see whether you thought about chunk size, metadata, and evaluation. If the only commit message is “final,” fix that before you apply. Five commits that show you changed the splitter after bad answers is better than a polished demo with no history.

• Helpful: a support-bot on public docs with citation links and a small eval CSV of questions you scored by hand.
• Helpful: a batch job that classifies tickets with a model and a fallback rule when confidence is low.
• Helpful: an internal copilot that drafts SQL you then review — say that humans still approve.
• Not helpful: “built ChatGPT clone” that is a UI on top of one API call.
• Not helpful: listing ten model names you typed into a playground once.

If your project used a company’s private data, do not paste secrets or customer text into a public repo. Describe the grain of the data (“40 internal how-to articles,” “anonymized ticket titles”) and the interface (Slack command, Next.js page, CLI). Interviewers care that you know what you were allowed to send to a vendor API. That sentence belongs on the resume because it shows you will not get the company banned.

Skills section: group tools like a human

Do not dump “Python, AI, ML, DL, NLP, ChatGPT, Bard, Gemini, Copilot, TensorFlow, PyTorch, Hugging Face, LangChain, LlamaIndex, RAG, LLM, Prompt Engineering, Generative AI.” That list is a keyword bag. Group: languages, app stack, model APIs, retrieval, eval, cloud. Put the posting’s words first only if you can talk for ten minutes. If you used OpenAI’s API and never trained a net, do not lead with PyTorch. If you fine-tuned a small classifier in scikit-learn, say that plainly. Classical ML still appears in fraud and ranking teams. It is not less impressive than a chatbot wrapper.

Soft skills belong in bullets, not as “excellent communication.” The communication that matters in this field is writing eval notes, documenting prompt versions, and telling a PM the model is wrong 12% of the time. If you did that, write it. “Wrote a one-page eval note on 50 golden questions after the bot started inventing refund policies” is a career skill. “Team player in an agile environment” is not.

Bullet formulas that survive a technical screen

Use verb + system + constraint + result. Weak: “Worked on generative AI solutions to enhance user experience.” Stronger: “Built a retrieval service in Python that answered questions from 80 public help articles; added source URLs after users complained about unsourced refund advice.” If you have no user count, say the corpus size, the latency budget, or the failure you fixed. Latency and cost are adult metrics in this market. “Cut average prompt tokens by trimming history to the last six turns” is better than a fake 300% engagement increase.

• Shipped a Next.js chat UI on Azure OpenAI with streaming tokens and a server-side prompt template stored in git.
• Chunked markdown docs with headings as metadata; switched from 2k to 800-token chunks after answers mixed two products.
• Logged prompts and completions with redaction for emails; used the log to find a jailbreak that asked for internal prices.
• Wrote 30 evaluation questions from real tickets and scored answers weekly so we knew when a prompt change got worse.

Notice none of those bullets say “leveraged cutting-edge LLMs.” They name a choice you made. Interviewers will pick one and ask why. If you cannot answer, delete the bullet. That is the whole quality bar.

Summary and title language

For a fresher: “Computer science graduate. Built a retrieval chatbot on public docs with Python and an OpenAI API; looking for a backend or applied-AI intern role.” For someone with two years of web work: “Full-stack engineer (TypeScript) who added an LLM search box to an existing Next.js app and owned prompt versions in git.” Do not write “AI enthusiast.” Do not write “prompt engineer” as a job title unless that was your actual title. Many hiring managers now treat “prompt engineer” on a fresher resume as a red flag for tutorial tourism.

Certifications and courses

A vendor certificate can sit in a short list if you also have a project. Alone, it is weak. Pair “Azure AI-900” with the thing you built on Azure. Pair “DeepLearning.AI short course” with the notebook you extended past the assignment. Do not list every LinkedIn learning badge. Two relevant courses beat twelve. If you are competing for ML research, courses will not substitute for math and papers. Be honest about which race you are in.

What to leave off

• Every foundation model you have chatted with.
• Claims that you “trained GPT.” You did not.
• Screenshots of ChatGPT conversations as a portfolio.
• A skills bar that scores yourself 90% in AI.
• Confidential customer transcripts in a public demo.

A one-page layout in MineResume

Use Classic or Compact so parsers keep Python, LangChain, and job titles as text. Header: email, city, GitHub, optional LinkedIn. Summary. Skills grouped. Experience or internships. Projects with two or three bullets each — this is where most of your proof lives if you are early career. Education. Certifications last. Keep the PDF to one page unless you have years of shipped systems. After export, search the PDF for “Python” and your GitHub URL. If they are not selectable, switch templates.

Tailor per posting. If they want Java and Bedrock, and you only have Python and OpenAI, apply only if the rest of the stack matches and you are ready to learn the vendor. Do not sprinkle “Bedrock” into Skills. Do rewrite “built RAG on markdown” into “document question-answering with embeddings and citations” if that is their phrase and it is still true.

Interview follow-through from the resume

Print your own bullets and write, in a notebook, why each technical choice happened. Why that chunk size. Why that model. What you would do with a private VPC and no third-party API. The resume’s job is to get that conversation. If a bullet cannot survive five follow-up questions, it is decoration. Cut it and spend the lines on a project you actually understand. That is how you look like an engineer in a market that is tired of AI wallpaper.

If you are still assembling proof, pick one corpus you care about — a college handbook, an open-source project’s docs, a public government FAQ — and build the boring pipeline: ingest, chunk, retrieve, answer, cite, log, eval. Put that repo in the header. Then write the resume around that one honest system instead of a cloud of tools. Hiring managers remember one working thing. They forget a list.

A week-by-week way to get interview-ready

If you are six months from graduating, do not try to learn every framework that appears in a Twitter thread. Week one and two: one language you can write without autocomplete — Python is enough. Week three: git, a README, tests for a function that splits text. Week four and five: call one model API from a small server, not a notebook only. Week six: add retrieval over a folder of files you are allowed to use. Week seven: write twenty questions and score the answers by hand. Week eight: deploy or at least record a two-minute walkthrough. That calendar produces a resume. A year of half-watched courses does not.

Working engineers adding AI to an existing product should spend their evenings on evaluation and safety, not on collecting model names. The market is full of people who can call an API. It is short on people who can say when the feature should not ship. Put one “we did not ship X because of Y” bullet if it is true. That sentence is senior even if your title is not.

India-specific application notes

Service-company portals often want a photo and a CV. Follow the portal. Keep a clean one-page MineResume PDF for product companies and startups that use Greenhouse or Ashby. Do not send a six-page academic CV into those. If a WITCH-style firm asks for every semester mark, that is their form, not a reason to clutter the startup PDF. Two files, two audiences. Same facts, different length.

Salary posts on LinkedIn for “GenAI” are often contractor theatre. Read whether the role is a six-month project. Contract work can still go on the resume with honest dates. Do not convert a two-month contract into “AI architect, 2024–present.” Hiring managers in Bengaluru and Hyderabad talk to each other. Dates that do not match your EPFO or your GitHub graph get noticed more than you think.

Sample one-page outline you can copy into the builder

• Header: name, email, phone, city, GitHub, LinkedIn.
• Summary: two sentences, job family plus one shipped AI feature or project.
• Skills: four groups, not twenty logos.
• Experience: two jobs or internships, three bullets each, one about reliability or eval if you have it.
• Projects: one or two, public if possible, failure case mentioned.
• Education and at most two certs.

When you paste this into MineResume, write the project bullets first. The summary is last. Export, then apply to roles where at least half the must-have tools are real for you. Volume applications with a fake stack waste their time and yours. The in-demand title is not a costume. It is a set of scars you can talk about on a call.

A longer look at evaluation, because that is the job now

In 2024 people demoed chat UIs. In 2026 the interviews poke at evaluation. If your resume cannot mention a golden set, a thumbs-down review, or a time a prompt change got worse, you look like a wrapper. You do not need a fancy harness. A spreadsheet of questions, expected points, and a weekly score is evaluation. Write that you scored it. Write that you stopped a release. That is senior behavior on a junior page, and it is still too rare.

Talk about latency and money in plain numbers if you have them: seconds to first token, dollars per day in a student account, a cap you set. If you do not have them, say you watched the vendor dashboard. Curiosity about cost is a hiring signal because someone on the team is about to get a bill shock. Be the person who noticed.

Safety is not a separate religion. It is product. Prompt injection, data sent to a vendor, users asking the bot to insult a coworker, medical or legal advice your bot should refuse. One bullet on a refusal or a filter is enough. Do not copy a policy PDF into the resume. Show a choice you made in the system.

If you are applying outside your country, say work authorization in one line. If you need sponsorship, some AI startups will not do it and some GCCs will. Do not hide it in a paragraph of dreams. Put it where a recruiter can see it, then let the project do the rest of the talking. The market is international and still bureaucratic. Your PDF should be both: a good project, a clear status.