We build production AI systems: retrieval systems, multi-agent architectures, and real-time voice agents. Most of our work starts after the prototype, at the point where a demo has to survive real users, real load, and someone being paged at 3am.
Voice agents. Built on LiveKit, where the hard problems are turn-taking, interruption handling, and holding round-trip latency low enough that the conversation feels like one. We have shipped voice assistants responding in under 500ms with tracing across the full pipeline.
Retrieval and multi-agent systems. LangChain and LangGraph, deployed on AWS, instrumented with Langfuse and Arize. What we get called in to fix is almost never the model. It is retrieval quality, evaluation, cost, and the absence of observability.
We publish components from our internal toolkit: skills, MCP servers, and agent tooling for problems that are hard to finish alone. Each component installs on its own, and each one is explicit about where it stops and what kind of engineering lies past that point.
| Repository | What it is |
|---|---|
hack2l-starter-kit |
A complete RAG application (FastAPI, Postgres with pgvector, LangGraph, Langfuse) plus a reviewer agent skeleton. Built as the target environment for our challenge at the Hack2L AI Agents Hackathon. |
Voice agent tooling is next.
We are based in São Paulo, Brazil, and take on AI engineering for teams in the United States and Europe. If you have an AI system that needs to reach production, or one that reached production and is not behaving, we are at vindler.solutions and you can book a call.
Engineers: we work with a small group of senior contractors, and we meet most of them through open source and through events like Hack2L. If you are building the same kinds of systems, the repositories above are a reasonable place to start a conversation.
