Q15 proof · Generative AI, agentic workflows and MCP

15. Generative AI, agentic AI and MCP

Answer as submitted

“Yes — RAG, embeddings, Llama-3 QLoRA fine-tuning, agentic workflows and MCP servers.”

Technical decision: expand. The submitted line lists techniques but not the engineering that makes them dependable. This page is the separate, detailed version: it shows retrieval-grounded answering, the honest choice between fine-tuning and prompting, typed tool-calling for agents, and the guardrails and evals that keep a probabilistic system inside its bounds.

Expanded answer

Yes — this is active practice, not theory. I build RAG pipelines (chunking strategy, embeddings, vector search over pgvector/FAISS-style stores, citation-forced prompting, context-window and cost control), run fine-tuning only where it earns its keep (QLoRA adapters on Llama-3-class models for style and format, never to patch missing knowledge), wire agentic workflows through MCP with schema-typed tools and confirmation on anything that writes, and put guardrails around the model: PII redaction before the prompt, prompt-injection checks on retrieved text, tool allow-lists and strict output schemas. Every prompt or model change is measured against a golden set with regression checks, and cost and latency are tracked per request.

Decision	Typical implementation	Why it is done this way
Grounding	retrieve → prompt with context → cite the source	An answer the business can verify, instead of text from memory.
Retrieval	chunking with overlap, hybrid vector + keyword search	Puts the right passage in context and nothing else.
Fine-tuning	QLoRA adapter for style and format	Shapes output without pretending to add knowledge.
Agents	MCP tools with typed schemas	The model's reach is exactly the tool contract — no more.
Write actions	confirmation step + optimistic concurrency	Autonomy stays bounded where state actually changes.
Privacy & safety	PII redaction, injection checks, output schema	Guardrails around the model, not promises inside the prompt.
Quality gate	golden set + LLM-as-judge + regression	A prompt change is released with evidence, like any other change.

Live GenAI delivery workflow proof

The running component below changes its decision/result view through real event binding. Select a step to see the technique, the reasoning behind it, and the measurement that keeps it honest.

GENAI WORKFLOWRUNNING

Choose a GenAI step

Select a step to inspect the decision, the reasoning and the measured result.

$ embed(query) → vector search (top-k) → prompt with context → answer with citations

The model answers from retrieved context and cites it, instead of answering from memory. That is what makes an answer verifiable and auditable by the business.

top-8 of 12,400 chunks  ·  recall@8 0.94
6 of 6 claims resolve to a source

Why the system is built this way

Rules I apply. Retrieve before generating and force a citation; keep knowledge in the index rather than in the weights; expose capability to an agent only as schema-typed tools, with confirmation on write actions; minimise data before the prompt and validate whatever comes back; measure every change against a golden set; and track cost and latency per request. The model is a component with a probabilistic boundary — the surrounding engineering is what makes it usable in a business process.

Boundary for Q15: this page covers applied GenAI delivery — retrieval, adaptation, agents and evaluation. Model pre-training, GPU infrastructure sizing and provider-specific commercial terms are separate topics.

Source files

Select a file to inspect the proof with syntax highlighting and line numbers. The first file is the code running the component above; the other two are the production shapes of the same decisions.

Acceptance criteria

Distinguishes retrieval, fine-tuning, agentic tool use, guardrails and evaluation.
Requires a citation for every claim instead of answering from memory.
Shows an interactive proof of choosing a GenAI step and reading its measurement.
Keeps writes behind confirmation and output behind a strict schema.
Defers pre-training, GPU infrastructure sizing and commercial terms to a later topic.