ai · ai agents, llm, claude, aws bedrock, nestjs
AI agent development: agents that take real actions, with guardrails.
A chatbot answers questions. An agent checks availability, books the table and remembers that the guest doesn't eat meat. I built one that does this for restaurants on WhatsApp, and it's live. I can build one into your product the same way, with a confirmation before every action and a cost cap per user.
- Models
- Claude Sonnet, Claude Haiku
- Platform
- AWS Bedrock, EU inference profiles
- Backend
- NestJS, Postgres with vector search
- Engagement
- As a project or as advisory
- Working hours
- Mon–Fri 04:00–18:00 Buenos Aires 09:00–23:00 Berlin 03:00–17:00 New York
- Available from
- Q4 '26
What goes into an agent that works
- Tools: the model calls functions of your system, such as checking availability or creating a booking. The WhatsApp assistant has 37, switched on per customer.
- The right model for each step: a rule-based router sends bookings to Claude Sonnet, which makes fewer mistakes, and simple questions to Claude Haiku, which is faster and costs about a third.
- Answers from your own data: FAQs found by vector search in Postgres and re-ranked before the model sees them.
- Guardrails: the model only proposes an action, and it runs once the user confirms. Validated inputs, timeouts and a daily cost cap per user on top.
- Tests and tracing: end-to-end conversations against the real models in CI, and every run traced, so you can see why the agent did what it did.
In production: reservations on WhatsApp
Restaurants connect their existing number to WhatsApp Business. Guests write, send voice notes or tap buttons, and the agent checks availability, takes reservations and answers questions about the menu, allergens and events, in the guest’s language. It remembers dietary needs for the next visit. One deployment serves all restaurants, each with its own instructions, tone and cost limits. Read the case study.
Inference in the EU
The models run on AWS Bedrock with EU inference profiles, so requests are processed in EU regions, and personal data is scrubbed before it reaches error tracking. For a European product, that’s often the first question.
How we can work together
- App development: an agent built into your product, with a written scope, a demo every week and the code in your hands. The backend is usually NestJS.
- Advisory: a second opinion before you build, on models, costs and guardrails, or a review of an AI feature you already have.
Questions
Which models do you use?
For the WhatsApp assistant, Claude Sonnet and Claude Haiku on AWS Bedrock. Which model fits depends on the task and the budget. Routing each step to a strong or a cheaper model keeps costs down without risking the steps that matter.
How do you stop the agent from doing something wrong?
Actions with consequences only run after the user confirms them. On top of that come validated inputs, timeouts, cost limits and a fixed fallback message if something fails. The confirmation step also limits what a prompt injection can do.
Can the agent use our existing systems?
Yes, that's what the tools are for: each one wraps a function of your backend, such as checking availability or creating a booking.
What does it cost to run?
That depends on the traffic and the models. Sending simple turns to a smaller model that costs about a third is one of the biggest levers, and a cost cap per user keeps surprises out.