AI Engineer Resume
Skills & ATS Keywords

Plan on roughly 32 to 48 named providers, frameworks, patterns, and tools, grouped into 8 short rows (foundation models, retrieval, frameworks, prompts, agents, evals, guardrails, inference and cloud). Less than 28 reads light for an AIE past the first year on the job; above 52 starts reading like a tag cloud. Treat each entry as a claim you can back with a shipped LLM feature: the prompt you iterated on, the index you own, the eval that gates the release. If the bullet does not exist, the line is taking up real estate without earning it.

Set the block right after your Profile Summary and before Work Experience. Parsers and recruiters both work from the top of the file down, and a labelled block sitting high up gets its terms surfaced more reliably. For an AIE, split it into 8 rows (foundation models and provider SDKs, retrieval and vector stores, app frameworks, prompts and structured outputs, agents and tool use, evals and observability, guardrails and safety, inference and cloud) so the parser reads clear clusters rather than a long ribbon of commas.

Paste the JD into a scratch doc, mark every model name, vector store, framework, and pattern that repeats two or more times, then collapse the highlights into a 12 to 18 item working list. Hold that list up against your Skills rows and against your shipped-LLM bullets. Anything the posting repeats that you actually run but that is missing from the page needs to land in the right row, and one of your bullets has to show you using it in a real retrieval, prompt, agent, or eval workflow. Run the cleaned file through an ATS Checker to verify the parser is reading the tokens you expect.

Both pages share Python and a willingness to deal with GPUs, but the spine is different. An ML Engineer page is built on production model infrastructure: distributed training (DDP, FSDP), the serving runtime (Triton, TensorRT, vLLM as a host), feature stores, drift monitoring, throughput numbers on classification, ranking, and recommendation models. An AI Engineer page is built on LLM applications: retrieval pipelines, vector stores, prompt-engineering frameworks, agents and tool use, eval harnesses, guardrails, and the cost-and-latency budget for inference calls against hosted models. Where an MLE bullet says trained the ranking model on FSDP across 32 H100s, an AIE bullet says shipped the citation-grounded RAG flow with a verifier pass and a golden-set CI. Pick one lane and order your bullets so the matching nouns hit the first scan.

No, and overclaiming here is one of the faster ways to get cut in a loop. The center of an AIE role is shipping LLM products on hosted or self-hosted models with retrieval, prompts, agents, and evals around them. Fine-tuning (LoRA, QLoRA, DPO) shows up on senior pages where the candidate has actually run a labeled dataset through a tuning pipeline, measured a lift against a frozen baseline, and held the resulting adapter in production. If you have done that, list it. If you have only read the paper or played in a notebook, leave it off and lean on RAG, prompt programs, and eval rigor instead.

Whichever one your team actually ships on. The honest answer is that 2026 hiring panels read a mix: LangChain and LangGraph for agent and orchestration code, LlamaIndex for retrieval-heavy RAG, and plain OpenAI or Anthropic SDKs (often with Pydantic for structured outputs) for teams that prefer thin wrappers. Lead with the one your last shipped feature ran on, and tag the others as supporting context only if you have written real code with them. Listing all three without a bullet to back any of them up reads as buzzword bingo and the screener marks it down.

Four families of numbers carry most of the weight on an AIE resume. Quality: eval scores on a named golden set, citation-grounding lift, hallucination rate before and after, factuality pass rate. Reliability: tool-call success rate, agent task completion, regression incidents caught in CI. Cost: per-call cost, monthly inference spend, dollar savings from prompt caching or model routing. Latency: p50 and p95 first-token time, p95 retrieval time, end-to-end response time under load. A bullet that names the feature, names the eval, names the cost-per-call, and names the latency reads as a real AIE shipping production work. Vague claims like improved LLM quality or optimized prompts get parsed once and skipped on the human read.