AI/ML Engineering

Build production AI systems with modern patterns and tools.

Quick Reference

The 2026 AI Stack

DSPy: Programmatic Prompting

Manual prompts are dead. DSPy treats prompts as optimizable code:

python
1import dspy
2
3class QA(dspy.Signature):
4    """Answer questions with short factoid answers."""
5    question = dspy.InputField()
6    answer = dspy.OutputField(desc="1-5 words")
7
8# Create module
9qa = dspy.Predict(QA)
10
11# Use it
12result = qa(question="What is the capital of France?")
13print(result.answer)  # "Paris"

Optimize with real data:

python
1from dspy.teleprompt import BootstrapFewShot
2
3optimizer = BootstrapFewShot(metric=exact_match)
4optimized_qa = optimizer.compile(qa, trainset=train_data)

RAG Architecture (Production)

Query → Rewrite → Hybrid Retrieval → Rerank → Generate → Cite
         │              │                │
         v              v                v
    Query expansion  Dense + BM25   Cross-encoder

LlamaIndex + LangGraph Pattern:

python
1from llama_index.core import VectorStoreIndex
2from langgraph.graph import StateGraph
3
4# Data layer (LlamaIndex)
5index = VectorStoreIndex.from_documents(docs)
6query_engine = index.as_query_engine()
7
8# Control layer (LangGraph)
9def retrieve(state):
10    response = query_engine.query(state["question"])
11    return {"context": response.response, "sources": response.source_nodes}
12
13graph = StateGraph(State)
14graph.add_node("retrieve", retrieve)
15graph.add_node("generate", generate_answer)
16graph.add_edge("retrieve", "generate")

MCP Integration

Model Context Protocol is the standard for tool integration:

python
1from mcp import Server, Tool
2
3server = Server("my-tools")
4
5@server.tool()
6async def search_docs(query: str) -> str:
7    """Search the knowledge base."""
8    results = await vector_store.search(query)
9    return format_results(results)

Embeddings (2026)

Fine-Tuning with LoRA/QLoRA

python
1from peft import LoraConfig, get_peft_model
2
3config = LoraConfig(
4    r=16,
5    lora_alpha=32,
6    target_modules=["q_proj", "v_proj"],
7    lora_dropout=0.05,
8)
9
10model = get_peft_model(base_model, config)
11# Train on ~24GB VRAM (QLoRA on RTX 4090)

MLOps Pipeline

yaml
1# MLflow tracking
2mlflow.set_experiment("rag-v2")
3
4with mlflow.start_run():
5    mlflow.log_params({"chunk_size": 512, "model": "gpt-4"})
6    mlflow.log_metrics({"faithfulness": 0.92, "relevance": 0.88})
7    mlflow.log_artifact("prompts/qa.txt")

Evaluation with RAGAS

python
1from ragas import evaluate
2from ragas.metrics import faithfulness, answer_relevancy, context_precision
3
4results = evaluate(
5    dataset,
6    metrics=[faithfulness, answer_relevancy, context_precision],
7)
8print(results)  # {'faithfulness': 0.92, 'answer_relevancy': 0.88, ...}

Vector Database Selection

Agents

• ai-engineer - LLM integration, RAG, MCP, production AI
• mlops-engineer - Model deployment, monitoring, pipelines
• data-scientist - Analysis, modeling, experimentation
• ml-researcher - Cutting-edge architectures, paper implementation
• cv-engineer - Computer vision, VLMs, image processing

Deep Dives

• references/dspy-guide.md
• references/rag-patterns.md
• references/mcp-integration.md
• references/fine-tuning.md
• references/evaluation.md

Examples

• examples/rag-pipeline/
• examples/mcp-server/
• examples/dspy-optimization/