AI Portfolio Projects That Actually Get Interviews

Hiring managers reviewing AI-engineer portfolios see the same three projects endlessly: a ChatGPT-wrapper chatbot, a PDF-Q&A demo, a LangChain tutorial clone. What gets interviews is evidence of engineering judgment: evaluation, cost awareness, failure handling, and a deployed URL. This guide gives project picks across levels, the quality bar that separates portfolio from tutorial-copy, and how to present them.

The quality bar (apply to any project)

A portfolio project earns interview discussion when it has:

A deployed URL — running beats repo-only, every time

An eval set and a score — even 50 labeled cases with a measured baseline→improved delta (how)

A cost line — "$0.004/query, here's the breakdown" signals production thinking

A failure-modes section in the README — what breaks it, what you did about it

One real design trade-off documented — "chose pgvector over Pinecone because X" (the reasoning)

One project with all five beats five projects with none — interviewers probe depth, not count.

Tier 1: the foundation project (everyone needs one)

RAG over a corpus you genuinely know — your field's regulations, a game's rulebook, a niche hobby's documentation. Domain familiarity is what lets you *judge answer quality*, which is what makes the eval set real. Stack: raw SDK (no framework — be able to explain every line), pgvector, FastAPI streaming, deployed. Differentiators: hybrid search vs pure vector (measured), chunking experiments (measured), citation accuracy checks.

Tier 2: pick one that matches your target role

Document pipeline (data/backend roles): invoices/forms → vision extraction → validated structured data → database, with confidence-gating and an accuracy report on 100 real documents. Boring domain, deeply hireable skill.

Agent with an approval gate (product/full-stack): a tool-using agent (LangGraph) that does something consequential — files issues, drafts emails — with human-in-the-loop interrupts, state persistence, and a budget cap. Demonstrates the autonomy-engineering employers are nervous about.

Eval harness for a public task (ML-adjacent roles): take a task (SQL generation, summarization), build the eval set, benchmark 4 models × 3 prompts, publish the matrix with cost columns. Pure judgment signal; zero UI needed.

Classifier at volume (infra-minded): a webhook-fed classification service with queue, idempotency, batch processing, per-1K-items cost — and a comparison against a fine-tuned small model (when that wins).

Tier 3: the conversation-starter (optional, memorable)

Something with personality that still shows engineering: a persona-consistent game NPC with memory, a local-first assistant respecting privacy by architecture, a domain-specific research agent with grounded citations (Perplexity-API-style). These get remembered; pair with Tier 1-2 substance.

Presentation: the README is the product

Structure that works: *what it does (2 lines + screenshot/demo link) → architecture diagram → eval results table → cost analysis → failure modes & mitigations → what I'd build next*. Then write one technical post per project on the non-obvious thing you learned ("my chunking experiment results", "where my agent burned $30") — posts get found, repos don't (the become-an-AI-engineer roadmap sequences this into 90 days).

Interview prep per project: be ready for "what breaks it?", "why this stack?", "what did it cost?", and "how would it scale 100×?" — your README sections are literally the answers.

What to skip

Tutorial clones with the tutorial's dataset; anything you can't run live in an interview; framework showcases where you can't explain what the framework does underneath; and quantity — three deep projects maximum, archive the rest.

---

*Last updated: June 2026.*