Now

A snapshot of the last twelve months, September 2025 to September 2026. Updated a few times a year.

Shipped

  • Seus Direitos, from WhatsApp bot to chat platform. The year started with multi-agent prompt experiments on a WhatsApp legal-intake flow: lead summaries generated by an LLM, engagement and hand-off messages, A/B tests across prompt versions and models. It ended with a channel-agnostic API and gateway serving WhatsApp, web, and app, rolled out to all web users, with per-request model overrides and a legacy mode for staged migration.
  • Migration onto an internal agent platform. Moved the agent onto Jusbrasil’s agent platform and added what the product needed along the way: thread-title opt-out, per-workspace billing labels, reasoning-effort defaults per deployment, and a model catalog that let us A/B Gemini tiers in production.

Measured

  • Evaluation service in Go. Evaluator prompts, judge overrides per run, boolean judge metrics exported to Prometheus, spreadsheet exports for the legal team, and a Pub/Sub path that survives pod interruptions.
  • Synthetic users from the CLI. A command-line client that plays a user against the agent, records eval-run notes, and reports latency, tokens, and estimated cost per turn. It is also the fastest way to compare two agent configurations side by side.
  • Daily evals as a gate. Scheduled runs, idempotent to pod restarts, so a prompt or model change shows up in quality metrics the next morning.

Operated

  • Observability for agents. Tokens per turn, duration per tool, reasoning effort, and conversation depth as Prometheus metrics; Grafana dashboards per workspace and model; alerts on latency by endpoint class, error rates, and drops in engagement.
  • Postgres and Kubernetes. Indexes on the hot polling paths, HPA and PDB for the gateway and API, a short-TTL session validation cache, and ArgoCD auto-sync.

How I work

  • Spec-driven changes reviewed and archived next to the code, so the repository carries the “why” as well as the diff.
  • Repository instructions and skills for coding agents (review checklists, log triage, test fixing), used daily by the team and by me.
  • Weekly written summaries to the team, and code review as a first-class activity: most days I review more pull requests than I open.