AI agents that remember.
In production, not in a deck.

Juanships LLC is the US software firm of Juan Pablo Osorio. We design, build and run AI agents with a memory layer underneath: systems that learn from every conversation they survive. Two of ours are in production right now.

• 107,750 appointments booked by our WhatsApp agent for healthcare clinics (Appcitas, live since 2015)
• 25,232 leads captured and qualified from Instagram, Facebook and WhatsApp (WalioAI)
• Enterprise copilot at VTEX RAG and multi-agent orchestration for B2B e-commerce
• paddock our open-source control plane for teams running fleets of Claude Code accounts

How an engagement works

1. Diagnosis, free. You answer four questions about what your product stores. We tell you what it already knows and isn't using, whether or not we end up working together. It starts with the one question that decides everything: does the same person come back?

2. Memory and agent architecture audit. Two to three weeks, fixed price. We map what closes in your domain, design the episode schema, the forgetting curve and the trust boundaries, and hand your team an architecture they can build without us. Priced after the diagnosis: three years of history and three months are not the same job.

3. Build and run. Monthly retainer. We build it into production alongside your team: bounded nightly costs, trust boundaries enforced in SQL, kill switches, and evals that tell you it works.

Probably a good fit

• Your product has been live for years and the database is the most valuable thing you own.
• Customers keep asking the same questions and your system answers each one like it is the first time.
• You have an AI feature that works, and you can feel that it is a commodity.
• Something in your product failed quietly for weeks because nothing was technically broken.

Probably not

• You want a chatbot in the corner of your app. That is the least interesting thing you can do with AI, and plenty of people will build it for you.
• Your product launched last month. You need customers before you need memory.
• You want someone to pick a vector database for you. That is not the hard part and it is not where the value is.

The memory patterns we ship were built in our own products first, twice, and written up in public: how I taught my ten-year-old product to dream.