An independent engineering practice in São Paulo, working with teams in the US and LATAM. Agent systems, contract software development, and its own products, on one stack.
Multi-agent architectures with PydanticAI and LangGraph, Postgres-backed state, Go and Python service layers, and the part most teams skip — evaluation, so an improvement is a number rather than an impression. The model is rarely the hard part. Everything around it that has to still be running at 3am is the work.
Scoped engagement, or embedded in your team. Six to twelve weeks.
Product or system built from nothing to live: backend, frontend, mobile, infrastructure, whatever the project needs. Botline owns the delivery, not a slice of it, and hands back code your engineers can keep alive.
Dedicated squad for a period, or fixed scope with a date.
For when something already exists and it will not move: an AI system that is unreliable, an integration nobody understands anymore, or an expensive architectural decision that has to be made before it becomes debt. Botline assesses it, says what is broken, and what fixing it would take.
Two weeks, written findings. The cheapest way to start.
Botline builds and runs its own products. It is where new approaches get tried before they go near a client system, and the evidence that the practice ships end to end rather than advising from the side.
Below.
Botline is deliberately small. Mauro leads every engagement and writes code on every engagement; specialists are brought in per project when the work calls for it. You talk to the person building it, from the first email to handover.
No retainer that quietly renews. When there is nothing left worth building, the engagement ends.
Client engagements, products and internal tools — everything Botline has shipped, in one list.
ERP item classification
The team was shipping prompt changes with no way to judge them. Botline built the evaluation harness — macro-F1 as the primary metric, deterministic scoring, automated error analysis feeding a structured improvement loop. Every change after that was judged against a number instead of an impression.
On-premise enterprise environment
No data could leave the network, which ruled out every hosted option before the project started. Botline authored MCP servers exposing internal systems to LLM agents, with transport, model selection and deployment all following from that single constraint. The agents work against live internal systems and nothing crosses the boundary.
Sports technology · United States
The product needed computer vision that worked on real footage, not on a curated validation set. Botline owned the loop end to end on AWS SageMaker — annotation, detection, pose estimation, temporal action segmentation, production inference — including the call on whether a model was good enough to ship.
Patas em Dia — pet health, routine and cost tracking. In beta.
Cafeteca — a brewing journal for people who take coffee seriously.
I read everything that arrives. If you have a system that needs to run in production and it involves agents, models or the plumbing around them, that is the conversation I want.
Contact Us