80+ buildable AI engineering portfolio projects for 2026 — grouped by theme, each with difficulty, the skills it proves, and annotated reference resources.
Maintained by Landed — daily AI-native job matches, agent help with every application, and mock-interview prep.
This is the broad, browsable catalog — 80+ genuinely distinct, buildable AI engineering projects across 10 themes, each tagged with a difficulty, the exact skills it proves, and 1–3 typed, license-noted reference resources. Pick by your skill gap, not by vibe: find the theme where your portfolio is thinnest, ship 3–5 projects with real evals and a live URL, and write each README as a product spec. One production project with proper evals beats five tutorial clones — hiring managers scan for observability, error handling, eval rigor, and a live link in about 90 seconds.
⭐ Star this repo — it's the exhaustive companion to the curated shortlist, refreshed for 2026.
Note
Want just the top 12 that actually land jobs? This catalog is breadth. For the curated, opinionated shortlist of the highest-signal builds — the ones hiring managers scan for — see the sibling repo projects-to-land-an-ai-job →.
flowchart LR
GAP["🔍 Find your<br/>skill gap"] --> PICK["📋 Pick 3-5<br/>projects"]
PICK --> EVAL["📊 Add evals<br/>+ live URL"]
EVAL --> SPEC["📝 README as<br/>product spec"]
SPEC --> PIN["📌 Pin 6,<br/>get referred"]
style GAP fill:#6C2BD9,color:#fff
style EVAL fill:#6C2BD9,color:#fff
style PIN fill:#00A86B,color:#fff
| Theme | Projects | What it proves | Catalog |
|---|---|---|---|
| 🔎 RAG apps | 12 | Retrieval, chunking, re-ranking, grounded citations | catalog/rag-apps.md |
| 🤖 Agents & tool-use | 12 | Orchestration, tools, memory, guardrails | catalog/agents.md |
| 📊 Evals & LLMOps | 10 | Golden sets, LLM-as-judge, CI gates, tracing | catalog/evals-llmops.md |
| 🎛️ Fine-tuning & training | 8 | LoRA/QLoRA, DPO, distillation, eval-vs-base | catalog/fine-tuning.md |
| 🎨 Multimodal | 8 | STT/TTS, VLMs, real-time streaming | catalog/multimodal.md |
| 🧩 Structured extraction | 7 | Pydantic/Zod, validation, confidence, retry | catalog/structured-extraction.md |
| 🔬 LLM from-scratch & internals | 7 | Tokenizers, transformers, training loops, CUDA | catalog/llm-from-scratch.md |
| ⚙️ Production & serving | 8 | vLLM/llama.cpp, quantization, latency, cost | catalog/production-serving.md |
| 🧠 Prompt & DSPy | 6 | Metric-driven prompt optimization, context eng. | catalog/prompt-dspy.md |
| 🚀 GTM & AI-PM prototypes | 7 | GTM automation, prototyping, AI-PM builds | catalog/gtm-ai-pm.md |
Total: 85 projects. Also below: what wins in 2026 · best build-with-me hubs · what's new · FAQ.
Difficulty legend: 🟢 Beginner · 🟡 Intermediate · 🔴 Advanced Link types: 📄 paper · 📘 docs · 🎬 video · 🧑🏫 course · 🛠️ tool · 💻 repo
Hiring managers don't clone your repo — they scan it. This is what they look for, in order:
| Signal | Why it wins | Where to add it |
|---|---|---|
| Production RAG (hybrid + reranking + citations) | Most-cited must-have skill | RAG apps |
| Eval suite (faithfulness, precision, hallucination rate) | "Eval is the new system design" | Evals & LLMOps |
| Live deployment URL (Vercel / Modal / Railway) | Reviewers won't clone — they click | Production & serving |
| README as product spec (problem, arch diagram, eval numbers, cost) | Reads as a shipped product | every project |
| Observability + cost tracking (Langfuse / Phoenix) | Signals production maturity | Evals & LLMOps |
| Error handling (retry, backoff, cost caps) | Signals reliability | Agents |
| Quantitative results (X% lift, $Y/1k tokens, p50/p95) | Numbers beat claims | every project |
| Open-source PR / issue history | Real collaboration, not a tutorial clone | any framework above |
Warning
Red flags that sink a portfolio: "GPT-4 wrapper, no original work" · "no eval, just vibes" · "benchmark claims without numbers" · "five copy-pasted Streamlit demos." Depth on 4–6 evaluated projects beats a long list of clones.
Per-project polish checklist: problem statement · architecture diagram · clean-clone setup (Dockerfile) · demo (live URL or 90-second GIF) · eval table (3+ metrics) · cost analysis ($/1k tokens, p50/p95) · failure modes · pinned repo (max 6) · a short design-tradeoffs write-up.
Each theme file lists the full project set with a one-line "what you build," the skills it proves, a difficulty rating, and annotated + license-noted references.
🔎 RAG apps → full list (12)
The highest-signal theme. Ship at least one hybrid-retrieval RAG with an eval table and citations.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| Document Q&A RAG | Hybrid PDF/Slack/Notion Q&A with citations | chunking, BM25+dense, citations | 🟢 |
| Contextual-chunk RAG | Anthropic contextual retrieval, measured lift | contextual chunking, recall@k | 🟡 |
| Agentic RAG | Retrieve-decide-retrieve loop | routing, iterative retrieval | 🔴 |
| GraphRAG | Entity graph + community summaries | graph extraction, summarization | 🔴 |
| Corrective RAG | Relevance grading + hallucination checker | grading, verification, guardrails | 🟡 |
| Multi-modal RAG (ColPali) | Retrieve over page images, not OCR | visual retrieval, VLM answering | 🔴 |
| Re-ranking pipeline | Cross-encoder rerank + nDCG/MRR lift | two-stage retrieval, ranking | 🟡 |
| Hybrid search from scratch | BM25 + dense + RRF fusion | lexical/semantic fusion | 🟡 |
…and 4 more (query-rewriting, text-to-SQL, long-context benchmark, NotebookLM clone).
🤖 Agents & tool-use → full list (12)
Add retry, backoff, cost caps, and tracing and one agent outshines ten toy loops.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| Multi-agent research team | LangGraph supervisor → cited report | routing, shared state | 🟡 |
| ReAct agent (5+ tools) | Reasoning loop over real tools | tool schemas, recovery | 🟡 |
| MCP coding agent | Reads issues, edits code, opens PRs via MCP | MCP, sandboxed exec | 🔴 |
| Persistent-memory agent | MemGPT-style long-term memory | memory hierarchy, retrieval | 🔴 |
| Voice support agent | STT → tools → TTS with barge-in | streaming, latency budgets | 🔴 |
| Browser / computer-use agent | Drives a real browser to finish tasks | vision-grounded action | 🔴 |
…and 6 more (CrewAI trip planner, hierarchical supervisor, plan-execute, game agent, briefing agent, Reflexion self-improver).
📊 Evals & LLMOps → full list (10)
The most under-built theme relative to how much it's asked about. "Eval is the new system design."
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| RAGAS eval harness | Faithfulness/precision/relevance on a golden set | golden sets, RAG metrics | 🟡 |
| LLM-as-judge + spot-check | Rubric judge validated against humans | rubric design, agreement | 🟡 |
| CI eval gate | Fail a PR on quality regression | CI/CD, thresholds | 🟡 |
| Trace cost/latency dashboard | p50/p95, tokens, $/1k per route | tracing, cost accounting | 🟡 |
| Red-team eval suite | Jailbreak/injection attack-success rate | red-teaming, safety | 🔴 |
| Drift monitor | Online quality + input drift alerts | monitoring, drift detection | 🔴 |
…and 4 more (A/B significance, hallucination audit, golden-set builder, model leaderboard).
🎛️ Fine-tuning & training → full list (8)
Only counts if you eval the tuned model against the base and report the delta.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| LoRA fine-tune vs base | Tune small Llama, head-to-head eval | LoRA, base-vs-tuned eval | 🟡 |
| QLoRA on consumer GPU | 4-bit tuning that fits 16GB | quantization, memory budgeting | 🟡 |
| DPO preference tuning | Align on preferred/rejected pairs | DPO, win-rate eval | 🔴 |
| Embedding fine-tune for RAG | Boost recall@k on your corpus | contrastive training | 🟡 |
| Reasoning distillation | Distill CoT into a small student | distillation, reasoning eval | 🔴 |
| GRPO on verifiable rewards | RL tune with checkable rewards | RL fine-tuning, reward design | 🔴 |
…and 2 more (Axolotl/Llama-Factory SFT, synthetic-data generation + filtering).
🎨 Multimodal → full list (8)
Demos incredibly well live. One polished voice or vision app is a memorable artifact.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| Whisper + diarization | Transcribe + label speakers | ASR, diarization | 🟡 |
| Blog-to-podcast agent | Article → two-host audio episode | summarization, multi-voice TTS | 🟡 |
| CLIP image search | Search an image library by text | joint embeddings, vector search | 🟢 |
| Chart/receipt digitizer | Vision → validated JSON | vision extraction, schema | 🟡 |
| Real-time voice bot | WebRTC + STT + agent + TTS, sub-second | real-time media, turn-taking | 🔴 |
…and 3 more (LLaVA visual chat, MiniGPT-4 captioning/VQA, text-to-image + eval).
🧩 Structured extraction → full list (7)
Quietly the most employable skill — every enterprise AI feature depends on it.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| Resume parser | Strict schema + per-field confidence | Pydantic, validation, retry | 🟢 |
| Invoice extractor + HITL | Extract + route low-confidence to review | nested schemas, review UX | 🟡 |
| Text → KG triples → Neo4j | Extract triples, load a graph DB | relation extraction, Cypher | 🔴 |
| Schema-locked agent output | Force every agent reply into a schema | constrained decoding | 🟡 |
…and 3 more (entity disambiguation, calibrated classification, web-page → records).
🔬 LLM from-scratch & internals → full list (7)
Doesn't ship a product — buys interview credibility. Pair each with a short write-up.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| BPE tokenizer | Train merges, encode/decode | BPE, vocab construction | 🟢 |
| nanoGPT (124M) | Reproduce GPT-2-scale training | training loops, LR schedules | 🟡 |
| llm.c (C/CUDA) | GPT-2 at the metal | CUDA kernels, low-level autograd | 🔴 |
| Tiny Llama | RoPE, RMSNorm, SwiGLU, GQA | modern-LLM building blocks | 🔴 |
| KV-cache + spec decoding | Measure tokens/sec speedup | inference optimization | 🔴 |
…and 2 more (Raschka full LLM build, attention visualizer).
⚙️ Production & serving → full list (8)
Where "I built a demo" becomes "I can run this in production." Put the numbers in the README.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| vLLM OpenAI-compatible API | Self-host + benchmark throughput | paged-attention, batching | 🟡 |
| llama.cpp edge deploy | Quantized GGUF anywhere | GGUF quantization, tokens/sec | 🟡 |
| FastAPI + SSE streaming | Production-shaped chat backend | streaming, auth, rate limits | 🟡 |
| One-click deploy (Modal) | Give any project a live URL | serverless GPU, secrets | 🟡 |
| vLLM stack on K8s | Routing, autoscaling, observability | K8s, GPU scheduling | 🔴 |
| Semantic-cache gateway | Cache + budget caps + fallbacks | gateway routing, cost control | 🟡 |
…and 2 more (MLX on Apple Silicon, batch inference pipeline).
🧠 Prompt & DSPy → full list (6)
Stop hand-tuning strings; program and optimize prompts against a metric.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| DSPy optimized pipeline | Compile prompts against a metric | DSPy modules, optimization | 🟡 |
| Prompt optimizer vs baseline | Beat your best manual prompt | automated search, held-out eval | 🟡 |
| Context-engineering study | What to put in the window, measured | context selection, compression | 🟡 |
| Prompt-injection playground | Attacks + layered defenses, measured | injection, defense-in-depth | 🟡 |
…and 2 more (prompt versioning + registry, CoT vs self-consistency benchmark).
🚀 GTM & AI-PM prototypes → full list (7)
The most under-built theme relative to demand. Build one and you're differentiated by default.
| Project | What you build | Skills | Difficulty |
|---|---|---|---|
| Context-aware sales agent | Multi-channel agent with memory | conversational memory, GTM | 🟡 |
| Lead-scoring + enrichment | Enrich + score against an ICP | enrichment, ICP scoring | 🟡 |
| Brief → landing page | One-line brief → deployable page | structured + image + code gen | 🟡 |
| Brief → prototype (AI-PM) | Feature brief → clickable prototype | spec parsing, code gen | 🔴 |
| Meeting-notes → CRM agent | Call → action items → CRM writes | transcription, extraction | 🟡 |
…and 2 more (support-ticket classifier, competitor/pricing monitor).
Don't start from a blank file — these hubs have runnable references for almost every project above. License matters if you plan to redistribute: ship your own work MIT, and only adapt MIT/Apache/CC0 code (with attribution).
| Hub | Stars | License | Best for |
|---|---|---|---|
| 💻 Shubhamsaboo/awesome-llm-apps | 116k | Apache-2.0 | 100+ ready-to-run agent + RAG apps — the widest starter set |
| 💻 rasbt/LLMs-from-scratch | 97.6k | Apache-2.0 (code) | Build a GPT from scratch: pretraining → fine-tuning, chapter by chapter |
| 💻 patchy631/ai-engineering-hub | 36.2k | MIT | In-depth tutorials across RAG, agents, evals, MCP, and serving |
| 💻 ashishpatel26/500-AI-Agents-Projects | 33.2k | MIT | 500+ agent use cases with links to open-source implementations |
| 💻 NirDiamant/RAG_Techniques | 28.2k | The best RAG-technique notebooks — re-implement, don't redistribute | |
| 💻 alexeygrigorev/ai-engineering-field-guide | 4.2k | check repo | Hiring research: take-homes, assignments, what interviewers actually ask |
Warning
License caution: only RAG_Techniques here is non-commercial — study and re-implement its techniques in your own code rather than copying notebooks into a repo you publish. Everything else is MIT/Apache-2.0 and safe to fork with attribution.
- Launched — 85 projects across 10 themes, each with difficulty, proven-skills, and annotated references. Sibling to the curated projects-to-land-an-ai-job.
- New in 2026: MCP-powered coding agents, computer-use/browser agents, GRPO/RL fine-tuning on verifiable rewards, reasoning-model distillation (R1-style), and ColPali-style visual document RAG.
- "Eval is the new system design" — expanded the Evals & LLMOps theme (CI gates, drift monitors, red-team suites) because it's the most under-built vs. how much interviewers ask.
- Context engineering added to the Prompt & DSPy theme — the 2026 replacement for string-fiddling.
How many portfolio projects do I need to get an AI engineering job? Fewer than you think — quality over quantity. Ship 3–5 deeply-evaluated projects, each with a live URL and an eval table, and pin 6 max on your GitHub. One production RAG project with proper evals beats five tutorial clones. Hiring managers scan for depth (observability, error handling, real numbers), not count.
What are good beginner AI project ideas? Start 🟢: a document Q&A RAG with citations, a resume parser with confidence scores, a BPE tokenizer from scratch, a CLIP image search, or a support-ticket classifier. Each is a weekend build that proves one clear skill — then add an eval table to level it up.
What are the best AI agent project ideas for 2026? The differentiated ones: a multi-agent research team (LangGraph supervisor), an MCP-powered coding agent, a persistent-memory agent, or a computer-use/browser agent. What makes them stand out isn't the framework — it's retry/backoff, cost caps, guardrails, and tracing. Add those and you're past 90% of portfolios.
How do I make a project stand out? Four things: (1) a live URL — reviewers click, they don't clone; (2) an eval table with 3+ metrics and real numbers; (3) a README written as a product spec (problem → architecture diagram → eval numbers → cost); (4) cost/latency figures ($/1k tokens, p50/p95). Bonus: an open-source PR to a framework you used. See the 90-second scan.
Should I build breadth or depth? Depth. Use this catalog to find the theme where your portfolio is weakest, then go deep on 3–5 projects there. For the curated shortlist of the highest-signal builds, see projects-to-land-an-ai-job.
PRs and issues welcome — add a project, fix a broken link, or update a star count. Every project must be buildable, and every entry must include skills + difficulty + at least one annotated reference. See CONTRIBUTING.md.
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
- 🧪 projects-to-land-an-ai-job — portfolio projects that actually get you hired
- 🗺️ ai-product-engineer-roadmap — the AI product engineer roadmap
Maintained by Landed · No affiliation with the companies or projects named. Content MIT-licensed; referenced repos retain their own licenses.