How to build an AI portfolio that gets you hired

A good AI portfolio is two or three deep projects that prove what the job ad is really asking for: that you can put a model into production, measure whether it works, and ship. Few postings ask for Kaggle or GitHub; many ask for production, evals and prototypes.

We read the full text of 2,362 live AI job postings on Tokn, in Europe and North America, as of 27 October 2026. Kaggle appears in one, GitHub as evidence in 10. Below: which projects to build for your role, how to present them and what to skip. For the whole job search, see how to get a job in AI, step by step.

In short

• Only 3% of the AI postings we read explicitly ask for a work portfolio. About 40% mention production work and about 20% mention evals.
• What counts differs by role: production and evals for engineers, papers and open-source code for researchers, shipped prototypes for product managers.
• Build two or three deep projects with a README, a live demo, an eval and a short write-up. Skip Titanic, MNIST and plain chatbot wrappers.

What do employers actually ask for in an AI portfolio?

Few postings ask for a portfolio by name. Many describe work a portfolio can prove: production systems, evaluations, prototypes, shipped products.

Method: we fetched the full public text of 2,362 live Tokn postings from their companies' applicant tracking systems (Ashby, Greenhouse, Lever) on 27 October 2026 and counted postings matching hand-checked phrase rules. Counts are mentions, including responsibilities and company descriptions, not strict requirements.

OpenAI and Anthropic make up about a third of the sample; without them, the picture holds (production 40%, evals 21%, portfolio 4%). Companies often judge projects in interviews without writing it down, so a rare mention doesn't mean nobody looks.

The application form is part of your portfolio

In a sample of 224 Greenhouse application forms at 21 companies (87 of them at Anthropic, so read it as indicative), 92% asked for LinkedIn and 68% had a "Website" field. Only 8% had a GitHub field, 6% a Google Scholar or publications field and 3% a portfolio link. A personal page about your projects fits the field most forms have.

Open questions often act as a portfolio review. Elastic's Agentic AI Engineer form asked: "Describe a GenAI or agentic project you built independently, end-to-end. What problem did it solve, what tools/frameworks did you use, and what made it 'autonomous' rather than just conversational?" Anthropic's Software Engineer, Labs form asked: "Describe a project you built from zero-to-one. What happened to it?" Have your answer written before you apply.

What do hiring managers look for in AI projects?

Public work they can judge quickly, with evidence that you measured it and maintain it.

Postings advertised in October 2026 put it plainly. Perplexity asked for "a public track record we can read, such as GitHub, a technical blog, or a demo we can try." Supabase listed as a "STRONG SIGNAL" an "open-source agent framework, an MCP server, an evaluation harness, or writing on agent reliability that other practitioners cite."

Anthropic's own advice on its careers page is to put "interesting independent research", "a thoughtful blog post" or open-source contributions "at the top of your resume". Louis-François Bouchard, who hires AI engineers, writes that "evals are the unit tests of AI engineering" and that "a broken setup or dead link signals that you do not maintain what you ship."

The right AI portfolio projects for your target role

Prove what postings for your role mention most (both regions; roles grouped by title).

For contrast, half of the 72 designer postings asked for a portfolio. For every other role, it's something you choose to build, not a box to tick.

ML engineer

Production appears in 70% of ML engineer postings, and Python and PyTorch are the most tagged skills in both regions. Show you can take a model past the notebook.

• Fine-tune a model in PyTorch, serve it with vLLM or FastAPI, and report latency and cost per request.
• Compare it with a simple baseline and show where it fails.

Expect to defend those choices: see the questions ML interviews actually ask.

AI / LLM engineer and forward deployed engineer

These postings mix production (60%), prototyping (44%) and evals (42%). See what an LLM engineer does for the job itself.

• Build an agent or RAG system on a real, messy corpus, with a labelled eval set, a failure analysis and a cost and latency table.
• For forward deployed engineer roles, build a working prototype for a real organisation's workflow, deploy it and write up the constraint you hit.

GitLab's Senior AI Engineer posting said "what matters is that you can deliver production-quality work independently."

Data scientist

Data scientist postings name almost no portfolio signal, so your projects carry more of the argument, not less.

• Answer an experiment or causal question with SQL and Python on data you collected, and state the decision it would change.
• For analytics-heavy roles in Europe, use dbt or BigQuery, which European postings tag.

Research scientist or engineer

Here papers count: 43% of research postings mention publications, and 24% of European ones (n=54) open-source contributions.

• Reproduce a recent paper, open-source the code and write up what didn't replicate.
• Contribute to projects such as vLLM, SGLang or Transformers.

AI product manager

PM postings ask for a portfolio almost never (1 of 119), but mention "shipped" and "prototype" in more than one in five. Mistral AI's Enterprise Controls PM posting wanted "technical depth to prototype, hack, or dive into code when needed." For the role itself, see what an AI product manager does.

• Build a working prototype with an AI coding tool, a one-page spec, an eval rubric and a short read-out of what users did with it.

How to present each project

Each flagship needs four parts: a README, a demo, an eval and a write-up.

The README

The first screen gives the problem, the result in numbers, the demo link and a one-command way to run it. Then say what you chose not to build, and why.

A demo people can try

Perplexity asked for "a demo we can try". Host it somewhere free and keep it running: a dead link is worse than none.

The eval

Bouchard's warning applies: "a demo that worked on one cherry-picked example tells an interviewer almost nothing." Build a small ground-truth set, compare a baseline with your approach, list failure cases and report cost and latency.

The write-up

Write one post per flagship: the problem, what you tried, what failed, the numbers. Link it from the top of your CV, as Anthropic suggests. And be ready to explain every design choice without notes.

What's overdone, and what to do instead

Tutorial datasets and plain chatbot wrappers prove you followed a tutorial. Keep the topic, but add what real work needs.

Elastic's question, what made your project "autonomous" rather than just conversational, is a good test for any LLM project. Certificates don't replace projects either; we cover which AI certifications are worth it separately.

Open source, papers and Kaggle: do they count?

Open source and papers count, mostly for research and infrastructure roles. Kaggle barely registers in postings.

Open-source contributions appear in 8% of European and 5% of North American postings, led by Mistral AI, Together AI, Anthropic and Scale AI. GitHub's Octoverse 2025 found that six of the ten fastest-growing open-source projects by contributors were AI infrastructure, including vLLM and SGLang. It also found about 81.5% of contributions happen in private repositories: most people's best work is invisible, so build public evidence on purpose.

Publications matter for research (43% of AI research jobs postings) and much less elsewhere: 13% of ML engineer postings and 5% of PM postings. Kaggle is fine for learning, but as your main exhibit it answers a question almost no posting asks.

Next steps

Give one flagship project a month: a week to define the problem, one for a baseline, one for evals and deployment, one for the write-up. Then apply with it at the top of your CV.

Browse AI engineer jobs or entry-level AI jobs to see what your target postings mention, and set up a job alert for your role.

FAQ

Do I need a portfolio to get an AI job?

Rarely as a written requirement outside design: on Tokn, as of 27 October 2026, 3% of live AI postings explicitly asked for one. But about four in ten mention production work and one in five evals. Without the job title, a portfolio is how you prove those.

Is a GitHub portfolio enough for AI jobs?

Rarely on its own. In our sample of 224 Greenhouse forms (skewed towards Anthropic), 8% had a GitHub field and 68% a "Website" field. Pin your best repositories, and add a short page that tells the story and the results.

How many projects should an AI portfolio have?

Two or three deep projects, one of them a flagship with a demo, an eval and a write-up.

Are Kaggle competitions worth it for an AI job?

For learning, yes. As evidence, they're weak: Kaggle appeared in 1 of the 2,362 postings we read. If you compete, deploy the model and write it up.

Can I use AI coding tools to build my portfolio projects?

Yes, and several application forms ask how you use them, but you must be able to explain every choice. Anthropic's guidance for candidates says: "We want to see your actual experience and how you think—not AI-generated responses."