12 deep, buildable AI/LLM portfolio projects that actually get people hired in 2026 — each a full brief with milestones, evals, and a deploy path.
Maintained by Landed — daily AI-native job matches, agent help with every application, and mock-interview prep.
The single highest-leverage thing for breaking into AI roles is a portfolio of shipped, measured projects. But quantity is a trap: one production project with proper evals beats five tutorial notebooks. This repo is the curated 12 — each a deep brief you can actually build, with the two layers everyone else skips: an eval report and a live deployment. That's the difference between "GPT wrapper" and "hire this person."
⭐ Star this repo — it's the quality-over-quantity companion to the broad 60+ catalog. Pick one, ship it, and use it as your referral conversation-starter.
A reviewer spends ~90 seconds on your repo. In those seconds they scan for four signals:
| ✅ The signal that wins | ❌ The red flag that loses |
|---|---|
| README as a product spec (problem, arch diagram, results) | 100 toy notebooks, no through-line |
| An eval report (named metrics + numbers + a target) | "it works well" — no evals, just vibes |
| A live deployment URL (reviewers don't clone) | runs only on your laptop, 12 undocumented steps |
| Cost & latency ($/1k tok, p50/p95) | benchmark claims with no numbers |
Build 3–4 of these deeply — not 15 shallowly — and present them per the presentation checklist.
flowchart LR
B["🔨 Build<br/>(in stages)"] --> E["📊 Eval<br/>(named metrics)"]
E --> D["🚀 Deploy<br/>(live URL)"]
D --> W["✍️ Write up<br/>(README + Loom)"]
W --> R["🤝 Get referred<br/>(project = the hook)"]
style E fill:#6C2BD9,color:#fff
style D fill:#6C2BD9,color:#fff
style R fill:#00A86B,color:#fff
Difficulty: 🟢 beginner · 🟡 intermediate · 🔴 advanced. Every brief includes Problem · Who it impresses · Skills · Stack · Milestones · Evals · Deploy · Stretch · Reference resources.
| # | Project | Skills it proves | Difficulty | Core stack | Brief |
|---|---|---|---|---|---|
| 1 | Chat with your docs (production RAG) | hybrid retrieval, reranking, citations, RAG evals, observability | 🔴 | LlamaIndex · pgvector · BM25 · reranker · RAGAS · Langfuse | → |
| 2 | Agentic workflow | tool design, orchestration, recovery, trajectory eval | 🔴 | LangGraph · tools · Langfuse | → |
| 3 | Production eval suite | golden sets, LLM-judge calibration, eval-gated CI | 🟡 | RAGAS · Phoenix · GitHub Actions | → |
| 4 | MCP "second brain" server | MCP, tool design, trust-boundary security | 🔴 | MCP SDK · pgvector · Claude Desktop | → |
| 5 | Real-time voice agent | streaming, STT/TTS, latency engineering | 🔴 | Pipecat/LiveKit · Deepgram · ElevenLabs | → |
| 6 | Structured extraction pipeline | structured output, schema design, validation/repair | 🟡 | Instructor/Outlines · Pydantic · Postgres | → |
| 7 | Fine-tuned specialist model | dataset curation, LoRA/QLoRA, eval vs base | 🔴 | Unsloth · PEFT · vLLM | → |
| 8 | LLM from scratch | tokenization, attention, pretraining internals | 🔴 | PyTorch · nanoGPT · minbpe | → |
| 9 | Semantic + hybrid search engine | ANN, BM25, RRF fusion, reranking, IR evals | 🟡 | Qdrant · bge · reranker · FastAPI | → |
| 10 | Guardrails & moderation | injection defense, PII, red-teaming | 🟡 | Guardrails-AI · Presidio · garak | → |
| 11 | Multimodal document AI | vision LLMs, multimodal RAG, chart/table QA | 🟡 | ColPali · VLM · docling · Instructor | → |
| 12 | LLMOps cost & latency dashboard | tracing, cost attribution, caching, alerting | 🟡 | Langfuse · OpenTelemetry · Redis | → |
| ＋ | GTM / AI-PM copilot (bonus) | applied AI, business value, real users | 🟡 | Next.js · Vercel AI SDK · Postgres | → |
Tip
Not sure which to pick? For AI/RAG Engineer loops, build #1 + #3. For AI Agent Engineer, #2 + #4. For LLM Engineer / Applied Scientist, #7 + #8. For AI Product Engineer / GTM, the bonus copilot + #6. See the FAQ.
- What makes a project job-winning
- The 12 projects (table)
- Project teasers — one paragraph each
- How to present your project — the polish checklist
- What's new (2026-07)
- FAQ
- Contributing
- The Landed family
1. Chat with your docs — production RAG 🔴 The flagship. Make an LLM answer questions about a real corpus with clickable citations, and prove the answers are grounded. You win by doing what tutorials skip: hybrid (dense + BM25) retrieval, a reranker, a "not in the docs" refusal, and a real eval table (recall@20, faithfulness, cost). The single most-cited must-have for RAG loops.
2. Agentic workflow 🔴 An agent that plans, calls ≥3 tools, recovers from failure, and finishes a real task — with a trajectory eval and pass^k reliability. The differentiator is retry budgets, structured tool errors, and a real verifier, not "the LLM called a function."
3. Production eval suite 🟡 The project that says "senior" because almost nobody builds it. A golden set from error analysis, a calibrated LLM-judge (Cohen's κ vs humans), an A/B test with significance, and a CI gate that fails a PR on regression. Directly answers "how do you know it works?"
4. MCP "second brain" server 🔴 Expose your notes/data as a Model Context Protocol server any AI client can use. Proves you understand tools as a trust boundary — tool-poisoning defense, version pinning, least-privilege — which is right at the 2026 frontier.
5. Real-time voice agent 🔴 A voice agent you can talk to: STT → LLM+tools → TTS, streaming, sub-second turn-taking and barge-in. The most impressive live demo in the repo, and it proves you can wrangle latency and streaming end to end.
6. Structured extraction pipeline 🟡 Turn messy invoices/resumes/receipts into validated, schema-locked JSON with confidence, validation, and repair. The pragmatic anti-hallucination flex that looks like real production work — because it is.
7. Fine-tuned specialist model 🔴 LoRA/QLoRA-tune a small model to beat a frontier API on one narrow task — cheaper and faster — and prove it head-to-head. The clearest "I can train, not just call APIs" signal for LLM-engineering loops.
8. LLM from scratch 🔴 Build a GPT — BPE tokenizer, attention, training loop — from first principles, then reproduce a small model. Makes every transformers-internals interview question trivial. A knowledge signal for LLM Engineer / Applied Scientist loops.
9. Semantic + hybrid search engine 🟡 RAG's retrieval half, isolated and done right: dense + BM25, RRF fusion, reranking, and a baseline-comparison table (recall@k, MRR, nDCG). The purest proof of the most in-demand RAG skill.
10. Guardrails & moderation 🟡 Wrap an LLM app in a safety layer — injection defense, PII redaction, jailbreak resistance — and red-team your own defenses with an attack-success-rate table. Safety maturity is rare in portfolios and disproportionately valued.
11. Multimodal document AI 🟡 An app that sees: answer questions from charts, tables, and scanned PDFs a vision LLM, with ColPali-style multimodal retrieval. The winning demo is a question whose answer is only in an image — text-only RAG returns nothing; yours nails it.
12. LLMOps cost & latency dashboard 🟡 Trace every LLM call — tokens, p50/p95 latency, cost, cache hits — in a dashboard with alerts, then use it to cut cost/latency and show the before/after. The production-mindset signal that separates "I built a demo" from "I run this."
＋. GTM / AI-PM copilot 🟡 (bonus) A business-facing build — lead enrichment + scoring + personalized outbound, ticket triage, or a 0→1 micro-app with 10 real users. Under-represented in portfolios, which makes it a differentiator for AI Product Engineer and GTM loops.
The build is half the work; presentation is the other half — and it's the half most people skip. The full guide is in presentation-checklist.md. The short version:
- README as a product spec — problem, a 90-sec demo GIF, an architecture diagram, a results table, "run it" in one command, "what I'd do next."
- Live URL — Vercel / Modal / Railway / HF Spaces. Put it at the top. Reviewers don't clone.
- Eval report — an
evals/folder, 3+ metrics with numbers and a target, a before/after, and a calibrated judge (say the κ). - Cost & latency — $/1k tokens, p50/p95, and the one thing you did to cut them.
- A writeup / Loom on the key tradeoff — this is the artifact you drop into a referral message.
- Pin max 6 repos on your GitHub profile. Curation is a signal.
- Rehearse the 90-second walkthrough and "how do you know it works?"
Warning
The fastest way to lose a loop with a good project is to not be able to explain your own eval. Build the eval you can defend — a copied RAGAS notebook you can't explain works against you.
- 2026-07 — Initial release: 12 curated deep briefs (+1 GTM/AI-PM bonus), the presentation checklist, and the contributor quality spec.
- Dated for 2026 — reflects the current signal set: eval is the new system design, MCP as the interop standard, agentic + trajectory eval, contextual retrieval, LoRA/QLoRA, real-time voice, cost/latency observability as table stakes.
- Cross-linked the broad 60+ catalog sibling throughout — this repo is the curated companion.
How many projects do I need to land an AI job? Three to four — built deeply, with evals and a live URL — not fifteen shallow ones. One production RAG project with a proper eval report beats five tutorial notebooks. Reviewers scan for depth and production signals in ~90 seconds; a curated, well-presented few wins. Pick from the table, and present them per the checklist.
What makes an AI project impressive in 2026? Four signals, in order: (1) a README that reads like a product spec, (2) an eval report with named metrics and numbers, (3) a live deployment URL, and (4) cost & latency figures. Bonus: observability, error handling (retry/backoff/cost caps), and quantitative before/after results. The red flags are the inverse — no evals ("just vibes"), no deploy, benchmark claims without numbers, and copy-pasted Streamlit demos.
Do I need to deploy it?
Yes — a live, clickable URL is one of the most-cited must-haves, because reviewers won't clone your repo. Use Vercel, Modal, Railway, or HF Spaces, ship a .env.example + Dockerfile, and put a rate-limit + cost-cap on the public URL. If the model must stay private, ship a recorded demo plus a one-command local run — never make the reviewer fight your setup.
RAG vs agent vs fine-tuning project — which should I build? Match it to the role. AI / RAG Engineer → production RAG + an eval suite. AI Agent Engineer → agentic workflow + MCP server. LLM Engineer / Applied Scientist → fine-tune + LLM from scratch. AI Product Engineer / GTM → the copilot + structured extraction. If in doubt, build the production RAG with evals — it's the highest-signal flagship across the most roles.
Do I need an ML PhD or to train models from scratch? No. Most AI Engineer / AI Product Engineer roles want people who ship with LLMs (RAG, agents, evals) — not researchers. LLM-from-scratch and fine-tuning are edges for LLM-Engineer / research-adjacent loops specifically; for most jobs, a deeply-evaled RAG or agent project wins.
Where do I get more project ideas?
This repo is the curated 12. For a broad menu — 60+ projects across RAG, agents, evals, fine-tuning, multimodal, serving, and GTM — see the sibling ai-engineer-portfolio-projects.
PRs and issues welcome — but this is a curated repo, so every brief clears a quality bar (it must prove a distinct skill and ship with evals + a deploy path). See CONTRIBUTING.md and use the Suggest a project issue template.
Part of the Landed AI-native job-search family:
- 🧭 awesome-ai-native-jobs — the umbrella that maps the whole AI-native job landscape
- 🔥 whos-hiring-in-ai — real hiring posts from founders on X, sorted by role
- 💸 recently-funded-ai-startups-hiring — fresh-capital startups staffing up now
- 🚀 ai-engineer-jobs — 300 live AI engineer roles, auto-updated
- 🤝 forward-deployed-engineer-jobs — FDE & customer-facing engineering
- 📈 gtm-engineer-jobs — GTM engineering roles
- 🎓 ai-fellowships-and-residencies — 75 fellowships, residencies & programs
- 📘 ai-interview-guides — 33 company interview guides
- ❓ ai-interview-questions — 331 real interview questions with answers
- 🗺️ ai-product-engineer-roadmap — the AI product engineer roadmap
- 📦 ai-engineer-portfolio-projects — 80+ buildable portfolio projects
