What are the 8 core areas of the 160-concept AI Engineer Roadmap?

The 160-concept mastery matrix spans 8 areas with 20 concepts each: LLM Foundations (Tokenization, KV Cache, LoRA, Quantization), RAG & Retrieval (BM25, HNSW, Hybrid RRF, GraphRAG, ColPali), AI Agents & MCP (ReAct, LangGraph, Context Compaction, A2A), Evals & Observability (Golden Datasets, LLM-as-a-Judge, Ragas, OpenTelemetry), AI System Design (Model Routing, Prefix Caching, vLLM, PagedAttention), AI Security (OWASP LLM Top 10, Indirect Prompt Injection, Sandboxing), Deployment & LLMOps (CI Eval Gates, Canary Rollouts), and Production AI Backend (SSE Streaming, Rate Limiting, Idempotency Keys).

What is the difference between Generative AI, Agentic AI, and AI Agents?

Generative AI produces single-pass outputs (text, code, images) from prompts; Agentic AI runs bounded, rule-guided workflows combining LLMs with deterministic routers and typed tool validation; AI Agents autonomously plan, invoke tools/MCP servers in a cyclic loop, inspect observations, and adapt until stop conditions or human-in-the-loop gates are reached.

Which RAG architecture should AI engineers deploy first in production in 2026-2027?

Start with Hybrid RAG (dense vector search + BM25 lexical search fused via Reciprocal Rank Fusion + cross-encoder reranking). Upgrade to Multimodal/ColPali Vision RAG for complex PDFs and diagrams, GraphRAG for multi-hop cross-document synthesis, and Adaptive RAG routers when mixing trivial chat with heavy retrieval.

160-Concept AI Engineer Cheatsheet Matrix, 17 ML Algorithms & 2027 RAG Blueprints

AREA 01 / 08 • LLM

LLM Foundations & Fine-Tuning

Tokenization, attention mechanics, structured outputs, reasoning models, LoRA & quantization.

0 / 20 Mastered

Subword BPE/tiktoken encoding maps text into integer IDs; explains character-counting & arithmetic quirks.

Interview Trap: Thinking models see characters or words instead of subword token IDs.

Decoder-only causal stack of Pre-Norm RMSNorm, Multi-Head/GQA Attention, and SwiGLU MLP blocks.

Interview Trap: Confusing encoder-only BERT with causal decoder-only GPT/Llama/Claude architectures.

Softmax(QK^T / sqrt(d_k))V scales dot products by sqrt(d_k) to prevent softmax saturation & vanishing gradients.

Interview Trap: Forgetting the 1/sqrt(d_k) variance proof or O(N^2) sequence complexity.

Effective usable context vs advertised max window; long prompts suffer "lost in the middle" attention decay.

Interview Trap: Assuming 1M context replaces RAG without hurting latency, cost, or needle recall.

Caches past Key/Value tensors during autoregressive decode so O(N^2) work drops to O(N) per step.

Interview Trap: Ignoring KV cache VRAM growth across batch size x layers x heads x seq_len.

Temperature scales logits before softmax; top-p (nucleus) truncates cumulative probability mass.

Interview Trap: Assuming temperature=0 guarantees 100% determinism across distributed GPU FP16 kernels.

Top-priority instruction hierarchy defining persona, XML boundaries, schemas, and negative constraints.

Interview Trap: Mixing untrusted user text with system directives without XML/tag isolation.

3-5 diverse edge-case input/output examples activate in-context patterns and cut schema errors by >90%.

Interview Trap: Adding 20 repetitive easy examples that bloat prefill tokens without covering edge cases.

Emitting intermediate reasoning tokens allocates inference-time compute before committing to final tokens.

Interview Trap: Expecting zero-shot single-token answers on multi-step math or code diff tasks.

Test-time compute models (o-series, DeepSeek-R1, Claude thinking) trained via RL on verifiable reasoning.

Interview Trap: Routing simple classification or boilerplate tasks to expensive high-latency reasoning models.

Constrained decoding (JSON Schema / CFG / Pydantic) guarantees 100% syntactically valid payloads.

Interview Trap: Prompting "return JSON" without constrained decoding or Pydantic retry validation.

Reuses KV states of identical long system prefixes and tool schemas for 50-90% input cost & TTFT savings.

Interview Trap: Putting dynamic timestamps or user IDs at the very top of the prompt, busting the prefix cache.

Asymmetric 3x-6x output-to-input token cost dictates concise responses, routing, and context trimming.

Interview Trap: Ignoring output token multipliers when estimating TCO for verbose agent traces.

Fluent fabrication caused by parametric memory guessing, weak retrieval, or missing abstention rules.

Interview Trap: Trying to prompt away hallucinations without retrieval grounding and explicit NOT_FOUND fallbacks.

Native vision, audio, and PDF page embeddings allow VLMs to read charts/tables without brittle OCR.

Interview Trap: Converting rich PDF diagrams into plain text OCR and losing visual layout relationships.

Freezes base weights W0 and trains low-rank B*A (B=0 init), mergeable for zero inference latency.

Interview Trap: Using fine-tuning to store rapidly changing facts instead of teaching format, style, or tool syntax.

4-bit NormalFloat (NF4) base weights + double quantization + paged optimizers fit 70B tuning on 48GB GPU.

Interview Trap: Forgetting that QLoRA dequantizes to BF16 during compute, trading speed for massive VRAM savings.

FP8, AWQ, GPTQ, and GGUF shrink weight memory 2x-4x so memory-bandwidth-bound decode runs faster.

Interview Trap: Quantizing sensitive activations/outliers naively without activation-aware scaling (AWQ).

Trains fast 3B-8B student models on verified frontier teacher reasoning traces and synthetic pairs.

Interview Trap: Distilling noisy unverified teacher outputs without golden-dataset decontamination.

Trade-off between frontier capability/zero-ops APIs and VPC data sovereignty + TCO crossover at scale.

Interview Trap: Self-hosting GPUs at 5k req/month where idle GPU baseline costs exceed API bills.

AI Engineer Visual Cheatsheets & 160-Concept Matrix

LLM Foundations & Fine-Tuning

Ready to test this matrix in real company interviews?