Most AI demos work once. My interest is in what comes after: making agentic systems observable, reproducible, and safe to change. So alongside the products I ship, I build the tooling that keeps them honest — conversation-level regression tests for agents, benchmarks whose every number traces back to a committed result file, and codebases that were test-driven from the first layer.

• 🔭   I build production AI systems — multi-agent LLM platforms, RAG, and ML — backed by solid web engineering
• 🧪   Currently deep in AI reliability: cross-turn agent invariants, deterministic replay, and measuring what people usually just assert
• 🌱   Leading FinRePAI · BitGov (AI credit-risk) and microservice development for RHIA (AI-powered HR platform) — both at IFCE's Innovation Hub (EMBRAPII)
• 💬   Ask me about LLM agents, agent testing, FastAPI/Django, RAG, observability, and shipping things that survive production
• 📫   Reach me at bessavagner.dev@gmail.com
• 📍   Crateús, CE, Brazil · Remote (UTC−3) — open to international freelance & roles
• 🗣️   Portuguese (native) · English (fluent) · Spanish (basic)

The thread through this work: if you can't replay it or measure it, you can't trust it in production.

ReplayGate — Conversation-level regression testing for multi-turn AI agents
Most agent-testing tools assert one turn at a time, which structurally misses the bugs that only exist between turns: the agent books before the user confirmed, or forgets a constraint set three turns back. ReplayGate records a real conversation once — every LLM call, tool call, and session transition — into a deterministic fixture, then replays it offline against a new model or prompt and asserts cross-turn invariants. It records at the application seam rather than the wire, so a fixture is readable JSON about what the agent actually did.
_{Python 3.12 · Pydantic v2 · DuckDB · Typer · OpenTelemetry}

stealthbench — A reproducible benchmark that puts numbers on browser-automation stealth
Instead of arguing about which automation setup is "stealthier," run them all through the same self-hosted detectors on the same machine and read the numbers. Every percentage traces to a committed results/*.json. Deliberately built so it can't be aimed at a third party: detectors run on localhost, results are numbers only, and it measures detectability without ever extracting data.
_{Python · Selenium · undetected-chromedriver · Camoufox · BotD · CreepJS}

RegWatch — "Google Alerts" for Brazil's Diário Oficial da União
Watches arbitrary terms on behalf of many clients, then ingests, matches, enriches and emails per-client regulatory digests — each mention carrying a one-line AI summary and a category. Built clean-room, TDD in every layer.
_{Django · PostgreSQL · LLM summarization}

_{🔒 Private or under NDA.}

Also maintaining Weberist (stealth web-automation framework), Aigents (provider-agnostic LLM adapters for Python), and Meta Agent (an AI that builds AI agents — 308 tests).

Languages

Backend & AI

Frontend

Data, Infra & Observability