# Interview Prep — AI/ML Engineer

```mermaid
mindmap
  root((AI/ML Engineer))
    Core ML/NLP
      Supervised vs Unsupervised
      Classification/NER/Sentiment
      NLTK, spaCy, HF Transformers
      Computer Vision (classification, detection, OCR)
    LLM Fine-tuning
      LoRA/QLoRA/PEFT
      RLHF/DPO
      Quantization & Distillation
      Fine-tune vs Prompt vs RAG
    RAG
      Chunking/Embeddings/Retrieval
      Hybrid search & re-ranking
      Multi-hop retrieval
      Vector DBs
        Pinecone
        Weaviate
        Qdrant
        pgvector
        FAISS
        Milvus
    Knowledge Graphs
      Neo4j / Graph RAG
      When graph beats vector search
    Agentic AI
      Tool calling & planning
      Multi-agent orchestration
      Human-in-the-loop
      Frameworks
        LangChain
        LangGraph
        LlamaIndex
        CrewAI
        AutoGen
    Multi-modal Systems
      Text + Image + Structured
      Distributed multimodal deployment
    Evaluation
      Offline/Online eval harness
      LLM-as-judge
      Hallucination detection
      Task-completion metrics
    MLOps and Serving
      vLLM / Triton
      Docker, CI/CD, Git
      Cloud - AWS/GCP/Azure
      K8s, Ray/Spark
      Kafka/Redis
    Observability
      LangSmith / Langfuse
      MLflow
      OpenTelemetry
      Arize Phoenix
    Governance and Security
      Data protection & access control
      Prompt-injection defenses
      Model risk & compliance
      Bias checks & auditability
```

---

## Checklist by Category

### 1. Core ML/NLP Foundations
- [ ] Supervised vs Unsupervised Learning
- [ ] Classification / Entity Extraction / Sentiment
- [ ] NLP libraries: NLTK, spaCy, HuggingFace Transformers
- [ ] CV pipelines: classification, detection, OCR, video analytics

### 2. LLM Fine-Tuning
- [ ] LoRA / QLoRA / PEFT
- [ ] RLHF / DPO (instruction tuning)
- [ ] Quantization & Distillation
- [ ] When to fine-tune vs prompt vs RAG (judgment question)

### 3. RAG (Retrieval-Augmented Generation)
- [ ] Ingestion → Chunking → Embeddings → Retrieval → Re-ranking
- [ ] Hybrid search, query expansion, multi-hop retrieval
- [ ] Grounded generation (reducing hallucination at generation time)
- [ ] Vector DBs: Pinecone / Weaviate / Qdrant / pgvector / FAISS / Milvus

### 4. Knowledge Graphs / Graph RAG
- [ ] Neo4j (or equivalent)
- [ ] When graph reasoning beats flat vector retrieval

### 5. Agentic AI
- [ ] Tool calling, planning, memory
- [ ] Multi-agent orchestration
- [ ] Human-in-the-loop checkpoints & guardrails
- [ ] Frameworks: LangChain / LangGraph / LlamaIndex / CrewAI / AutoGen
- [ ] Judgment: when NOT to use a framework

### 6. Multi-Modal / Multi-Model Systems
- [ ] Combining text + image + structured data in one pipeline
- [ ] Large-scale distributed multimodal training/inference

### 7. Evaluation & Hallucination
- [ ] Offline/online eval harnesses
- [ ] LLM-as-judge
- [ ] Hallucination detection & mitigation
- [ ] Task-completion measurement

### 8. Serving & Infra (MLOps)
- [ ] Inference servers: vLLM, Triton
- [ ] Docker, CI/CD, Git, REST APIs, SQL
- [ ] Cloud: AWS / GCP / Azure
- [ ] K8s, Ray/Spark (distributed compute)
- [ ] Kafka/Redis (streaming/caching)

### 9. Observability
- [ ] LangSmith / Langfuse / MLflow / OpenTelemetry / Arize Phoenix

### 10. Governance & Security
- [ ] Data protection, access control
- [ ] Prompt-injection defenses
- [ ] Model risk & compliance monitoring
- [ ] Auditability, bias checks

### 11. Core ML Stack
- [ ] Python (production-quality, tested code)
- [ ] PyTorch, Scikit-learn, pandas
- [ ] Regulated industry exposure (finance/healthcare/insurance) — bonus