Back to selected work

Conversa Labs and Maestro: a support platform with controlled AI actions

Engagement
Maestro · Conversa Labs
Period
Conversa Labs engagement · November 2025–present
Stack
  • Python
  • FastAPI
  • LangGraph
  • Langfuse
  • PostgreSQL
  • pgvector
  • Redis
  • Ruby on Rails
  • Vue
  • MCP
  • OAuth 2.1
  • Docker Swarm
MaestroConcept preview
  1. Done: Recover conversation state
  2. Done: Call an authorized tool
  3. Current step: Hand off to a person

The platform brings its own workflows and AI execution together. Maestro can hold a reply whose claimed action does not match tool execution, request human approval and pause when an operator takes over.

In 30 seconds

Problem

Customer support, business workflows and AI actions need the same account context, permissions and human control.

My part

I’m a partner and the main developer of Conversa Labs’ own platform modules and services. I built modules on Chatwoot’s open-source (MIT) core, without the Enterprise edition, and built Maestro as the AI agent service.

Result

The platform brings its own workflows and AI execution together. Maestro can hold a reply whose claimed action does not match tool execution, request human approval and pause when an operator takes over.

Technical decisions

  1. Build own modules on the MIT core, without the Enterprise edition.

    Why
    Shared account context connects the modules to support permissions and conversations.
    Trade-off
    Keeping the modules compatible requires continuing work when upstream changes.
  2. Check claimed actions against executed tools before sending.

    Why
    An agent reply needs to reflect the actions that actually ran; sensitive actions require human approval.
    Trade-off
    Checks may hold a reply, and human approval introduces a waiting state.
  3. Keep memory and cost attribution account-scoped.

    Why
    Deterministic checks can hold a reply before delivery. An optional sampled LLM judge evaluates afterward without changing it.
    Trade-off
    Semantic retrieval needs an embeddings index; evaluation falls back to rules when the optional judge is unavailable.

Operation

The Rails/Vue work includes CRM/Kanban, paid-order commission reconciliation, a WhatsApp Web channel, a visual flow builder, support routing, granular permissions and an OAuth 2.1 MCP server. Maestro uses Python, FastAPI, LangGraph, PostgreSQL/pgvector and Redis for conversation state, account/contact memory and hybrid search, with semantic retrieval when configured. I instrumented AI usage and cost estimates per account and conversation, including fallback, with Langfuse tracing when enabled. I built a contract-derived reference generator with private-term checks before publication, an Astro subscription checkout and trial integration, and configured cosign signing, SBOM generation and Docker Swarm delivery.

  • CRM/Kanban
  • Paid-order commission reconciliation
  • WhatsApp Web channel
  • Visual flow builder
  • Support routing
  • Granular roles and permissions
  • MCP server

Authorship

The MIT license applies to Chatwoot’s open-source core. My contribution covers the platform’s own modules, Maestro service and integrations, while upstream authorship remains with its contributors. I have worked with Conversa Labs since November 2025.

Discuss an engineering role

Open to on-site and hybrid opportunities throughout Spain, and remote opportunities in Europe.

contato@guilhermejansen.com

Guilherme Jansen · Senior Software Engineer · AI Engineer

Back to top