One person instead of three suppliers.
I map the process, build the agent, and ship the application it runs on.

• Process first. I came to this from industrial engineering, an MBA and years in financial consulting. I start from how the business actually works, not from the model.
• The AI layer. Agents, retrieval, and the part that takes the time — deciding what the agent isn't allowed to do.
• The application underneath. Agents need somewhere to live. I build that too, so there's no coordination tax and no handoffs.

11 tools: four retrieve from their legal corpus, three act, four manage its own internal state. It never does arithmetic, because it quotes fees — it calls a whitelisted evaluator instead. Without country, qualification and year, it asks rather than guessing at a legal answer.

The owner types an order into a chat. The system issues the invoice, records the expense, updates the CRM and feeds a dashboard. 22 tools, a separate read-only role for their accountant, and EU VAT validated against the official registry rather than a regex.

No menus, no forms, no flow. A conversation that long outgrows any context window, so memory is tiered: Redis for the live conversation, Postgres for what has to survive it. Answers are retrieved from the corpus rather than packed into the prompt.

An agent will lie to itself across turns if you only persist what it said.

A tool failed silently in one of my systems and the model announced success anyway. That announcement went into conversation history, and on the next turn the model read its own fiction as fact and acted on it. The fix was persisting the tool execution log alongside the assistant message, so the agent's memory holds what happened instead of what it claimed happened.

That failure mode is invisible until it isn't.

Tools I built because production kept needing them. All TypeScript, all with CI and versioned npm releases:

• ragproof — your RAG, put to proof: self-correcting answers and CI eval gates that block the merge when hallucination rate crosses your threshold.
• llm-quorum — route AI tasks to the right model CLI, with automatic fallback and cross-model consensus review.
• n8n-workflows-mcp — turn your deployed n8n workflows into tools your AI agents can call, without rewriting anything.
• castloop — multi-agent content pipeline that doesn't break in production: research → publish → analytics, every handoff judged and traced, human gate before publishing.
• instinctkit — survival kit for self-learning Claude Code setups: starter instincts plus guardrails against silent agent failures.
• listingkit — one property listing in, six branded marketing assets out.

Client work lives in private repositories. Happy to walk through architecture and decisions on a call — usually more useful than a repo tour anyway.

The forks on this profile are tools I use in my own workflow, not projects of mine.

🌐 nestormartinez.dev  ·  🧭 podervsfuerza.com  ·  💼 LinkedIn  ·  ✉️ info@podervsfuerza.com

_{Madrid, working remotely · Available for project and hourly work}