AI Portfolio Projects That Get You Hired (Mapped to Certifications)

By Kyle Tice · Updated October 19, 2026 · Facts checked against official exam pages

The AI projects that get interviews in 2026 show the same skills employers list in job posts: retrieval over real data, agents that call tools safely, evaluation you can measure, models that run in production, and documentation of risk. Five projects cover all of that: a RAG app over a niche corpus, an MCP agent with guardrails and an eval harness, a fine-tune with a proper evaluation, an MLOps pipeline, and a governance artifact. Each one below is mapped to the certification it prepares you for, so the same work counts twice.

You don’t need all five. Pick the two that match your role roadmap and do them well.

What makes a portfolio project count

Hiring managers look at dozens of chatbot demos. The ones that stand out share four traits:

1. A real, specific dataset. “Chat with my PDFs” is generic. “Answer questions over 3,000 county zoning ordinances, with citations” is a project.
2. Numbers. Retrieval hit rate, answer accuracy on a test set, latency, cost per query. Stack Overflow’s 2025 survey found 46% of developers distrust the accuracy of AI output, and evaluation results are how you answer that skepticism.
3. Failure handling. What happens when retrieval finds nothing, the model returns malformed JSON, or a tool call fails?
4. A write-up. A README that explains the problem, the architecture, the results and what you’d do next. Most reviewers read the README and skip the code.

Keep each repo runnable: a clear setup section, an .env.example, and a small sample dataset if the full one is too large or licensed. Never commit API keys.

Project 1: RAG over a niche corpus

What you build: a retrieval-augmented generation app over documents in a domain you know. Good options are public regulations, technical manuals, clinical guidelines, court opinions, a product’s support docs, or research papers in one subfield.

Minimum scope:

• Ingestion: parse, clean and chunk the documents. Try at least two chunking strategies and compare them.
• Retrieval: vector search plus keyword (hybrid) search, with a reranker if you can.
• Generation: answers with inline citations to the source chunk.
• Evaluation: a golden set of 50–100 questions with known answers and source passages. Report retrieval recall and answer accuracy for each configuration.
• A “can’t answer” path when retrieval confidence is low.

Stretch goals: metadata filtering (date, jurisdiction, document type), access control so users only retrieve documents they’re allowed to see, and a cost and latency table.

Certifications it reinforces:

Time: 2–4 weeks part-time. Cost: usually under $20 in API and embedding calls for a corpus of a few thousand documents, if you cache embeddings.

Project 2: Agent with MCP tools, guardrails and an eval harness

This is the most in-demand project type right now. AI engineering tops LinkedIn’s 2026 Skills on the Rise list, and agents with tool use are where much of that work is.

What you build: an agent that completes a multi-step task by calling tools you expose through your own MCP server. Examples: a research assistant that searches an API, reads results and files a summary; an ops agent that queries logs and opens tickets; a data agent that runs read-only SQL and builds a report.

Minimum scope:

• An MCP server with 3–5 tools, typed input schemas and clear tool descriptions, built against the current MCP spec.
• An agent loop that plans, calls tools, handles tool errors and stops cleanly.
• Guardrails:
  • Allow-listed tools only, with least-privilege credentials (read-only where possible).
  • Human approval before any action that writes, sends or deletes.
  • Input checks for prompt injection in retrieved or tool-returned content.
  • Output validation (structured output against a JSON schema) before results are used.
  • Limits on steps, tokens and spend per run.
• An eval harness: 20–50 scripted tasks with expected outcomes. Measure task success rate, number of tool calls, cost and latency. Add an LLM-as-judge check for open-ended outputs, and run the whole suite as a regression test whenever you change a prompt or model.
• Traces: log every tool call and model response so you can show a failed run and explain it.

Stretch goals: a red-team section showing an indirect prompt injection (a poisoned document that tries to make the agent call a tool it shouldn’t) and how your guardrails block it. This doubles as a security portfolio piece for CompTIA SecAI+.

Certifications it reinforces:

Free prep: the Hugging Face MCP Course and Agents Course, Anthropic Academy’s Intro to MCP and MCP Advanced Topics, and the Anthropic Cookbook’s tool-use examples.

Time: 3–5 weeks. Cost: $10–$50 in model calls, mostly from running the eval suite repeatedly. Use a cheaper model for development runs.

Project 3: Fine-tune and evaluate a small model

What you build: a fine-tuned open-weights model (small enough to train on one GPU with LoRA or QLoRA) for a narrow task, plus a fair comparison against prompting a base model.

Good tasks: classifying support tickets, extracting fields from domain documents into JSON, or rewriting text in a specific style. Pick something where you can score outputs automatically.

Minimum scope:

• A clean dataset with a held-out test split that you never train on. Document where the data came from and its license.
• A baseline: the base model with a good prompt, and a few-shot version.
• A fine-tune with tracked hyperparameters.
• An honest results table: accuracy (or F1, or exact match), latency and cost per 1,000 requests for baseline vs fine-tuned. If fine-tuning didn’t beat prompting, say so. That is a useful result and shows judgment.
• A model card (see Project 5) describing intended use and limitations.

Certifications it reinforces:

Time: 2–4 weeks. Cost: free on Kaggle or Colab GPUs for small models, or $10–$50 of rented GPU time.

Project 4: An end-to-end MLOps pipeline

What you build: a pipeline that takes a model from training to production and keeps it healthy. The model can be simple, such as a tabular classifier. The pipeline is the point.

Minimum scope:

• Versioned data and a reproducible training job.
• Experiment tracking (MLflow or the cloud equivalent) with a model registry.
• CI that runs tests and an evaluation gate: the new model only deploys if it beats the current one on a held-out set.
• Deployment to an endpoint (managed cloud service or a container).
• Monitoring for data drift and prediction quality, with an alert or automatic retraining trigger.
• For a GenAI version, swap the classifier for a prompt or RAG app and gate deployments on your eval suite from Project 1 or 2.

Certifications it reinforces:

Build it on the cloud you’re certifying in. Many exam questions describe exactly the decisions you’ll make.

Time: 3–5 weeks. Cost: $0–$50 if you stay within free tiers and tear resources down after each session. Set a billing alert first.

Project 5: A governance artifact

Governance work is documentation, and documents are easy to show. This is the right project for governance, legal, compliance and PM candidates, and a strong add-on for engineers.

What you build: a complete governance package for one realistic AI system. Use your own project from above, or a well-documented public use case such as a resume-screening tool (a high-risk use under Annex III of the EU AI Act).

Minimum scope:

• System description: purpose, users, data sources, model, human oversight.
• Risk classification under the EU AI Act, with reasoning.
• AI risk or impact assessment: risks to individuals (bias, privacy, safety), likelihood and severity, mitigations, and residual risk. Map each risk to NIST AI RMF functions (Govern, Map, Measure, Manage).
• Model card: intended use, out-of-scope uses, training data summary, evaluation results broken down by relevant groups, known limitations.
• Control mapping: a table linking your mitigations to ISO/IEC 42001 Annex A controls.
• Monitoring plan: what you measure after launch and who reviews it.

Publish it as a clean PDF or a docs page. If you base it on a real employer system, get permission or anonymize it fully.

Certifications it reinforces:

See the AI governance certifications guide for which credential to pair it with.

Time: 1–3 weeks. Cost: $0.

Which projects for which role

Presenting your work

• Pin the repos on GitHub and link them in your resume header, not just a footer.
• Lead each README with results: one sentence on the problem, one table of numbers, one architecture diagram.
• Record a 2–3 minute demo video. Many reviewers won’t run your code.
• Write one post per project explaining a decision you made and why. This gives interviewers something to ask about.
• Put the certification next to the project. “Passed AIP-C01; built a Bedrock RAG app with 91% answer accuracy on a 100-question test set” is far stronger than either alone. (Use your own real numbers.)

FAQ

Do I need a portfolio if I have certifications?

For technical roles, yes. Employers outside regulated industries consistently rank demonstrated work above certificates. Certifications help you get past filters, and the portfolio is what you discuss in the interview. See what employers want.

Can I use paid APIs, or should I stick to open-source models?

Either is fine. Using a hosted API (Claude, OpenAI, Gemini, Bedrock) is closer to how most companies build. Showing you can also run an open model locally is a plus. What matters most is that you measured the results.

How big should the dataset be?

Big enough that naive approaches fail. For RAG, a few hundred to a few thousand documents. For evals, 50–100 golden examples is a credible start. Quality and honest reporting matter more than size.

Is it okay to follow a tutorial?

Use tutorials to learn, but change the project substantially: a different dataset, an added eval suite, added guardrails. Reviewers recognize tutorial clones. Your write-up should explain choices the tutorial didn’t make for you.