Most people learning AI engineering collect definitions. Interviewers ask for mechanisms. Those are not the same skill. Wrong question: "what is RAG?" Right question: "why retrieve before you generate?" 12 concepts. 12 questions. Watch each one actually run 👇 🟢 FOUNDATIONS 01 Fundamentals · AI ⊃ ML ⊃ deep learning 02 Prompting · role, task, format, limits 03 Few-Shot · show 3 examples, it copies the 4th 🔵 RETRIEVAL 04 RAG · retrieve, then generate 05 Embeddings · "refund" becomes [0.12, -0.31…] 06 Vector DB · nearest-neighbour search, top-k 🟡 MODEL OPS 07 Inference · tokens out, one at a time 08 Evaluation · golden sets + LLM-as-judge 09 Fine-Tuning · adapt weights, not prompts 🟣 AGENTS & SAFETY 10 Tools · the model calls your API 11 Agents · plan → act → observe → repeat 12 Guardrails · filter input and output 👀 Look at card 09. Most fine-tuning diagrams show a whole network glowing. That's wrong for how anyone actually does it in 2026 — with LoRA, ~99% of the weights stay frozen and you train a thin adapter on top. That single detail is why fine-tuning is cheap now, and why "fine-tune vs RAG" isn't the budget question people still treat it as. And card 07: the tokens appear one at a time because that's literally what's happening. Autoregressive generation is the reason latency is measured in time-to-first-token, not total time. The rule: if you can't draw it moving, you can't explain it under pressure. Most people prep by reading definitions, then get asked "walk me through what happens." Send this to whoever is interviewing this month. 📸 Screenshot it — 12 concepts, 12 questions, one card. Follow @hackproduct — scary AI concepts, made shippable. ⚡ . . #AIengineering #LLM #RAG #AIagents #vectordatabase