AI Engineer — LLM Agents · RAG · MLOps
MSc in Computer Engineering · PhD Student @ UNICAMP
📍 Campinas, Brazil
I design and ship LLM agents that run in production — not demos. My work sits where agent orchestration meets real infrastructure: retrieval over private knowledge bases, tool calling against live systems (BigQuery, CRMs, WhatsApp), input guardrails, persistent memory, and full tracing so failures are debuggable instead of mysterious.
In parallel, I'm a PhD student in Computer Engineering at UNICAMP researching reinforcement learning for robot skill acquisition.
Agents & LLM frameworks
RAG, data & observability
Backend & MLOps
ML & Robotics
No hardcoded prompt: every GoHighLevel sub-account configures its own agent from the dashboard, and the runtime loads that config dynamically per incoming message — one service serving all tenants. Inbound webhooks return fast and enqueue; a Redis debounce buffer merges rapid-fire user messages into a single inference; async workers then run one full turn — agent factory, chat history, pgvector RAG retrieval and CRM tools (scheduling, tags, custom fields) — before replying.
FastAPI · LangChain · Supabase/pgvector · Redis · Next.js · Multi-tenant
Production support agent serving WhatsApp through Chatwoot. RAG over a private knowledge base (Pinecone), contextual web search (Tavily), persistent conversation memory (PostgreSQL), and human handoff when the agent should step aside. Ships with an 8-layer input guardrail and end-to-end Langfuse tracing plus LLM-as-a-Judge scoring.
LangChain · FastAPI · Pinecone · PostgreSQL · Chatwoot · Langfuse
Turns natural-language questions into valid BigQuery SQL, executes it, and answers in plain language. Built as a LangGraph state graph cycling between an agent node and a tool node until the answer is grounded in real query results. Runs on the public NYC Citi Bike dataset, served via Streamlit and containerized with Docker.
LangGraph · BigQuery · SQLAlchemy · Streamlit · Docker
Give it a company URL and three specialized agents hand work down a chain: a scraper crawls the site (Apify), a profiler infers pain points, tech stack and ideal customer, and a copywriter writes a personalized cold email. Non-conversational, fully orchestrated as a LangGraph pipeline.
LangGraph · Multi-agent orchestration · Apify
You speak, Whisper transcribes, a LangChain agent reasons over a pandas dataframe of São Paulo rental listings, and the answer comes back as speech via OpenAI TTS. Full voice-in / voice-out loop.
LangChain · Whisper · pandas · OpenAI TTS
Real-estate sales agent wired into Kommo CRM through webhooks: answers from a RAG policy base, searches listings in Google Sheets, and drives the sales funnel (moves stages, updates custom fields) via tool calling. Swaps between OpenAI and a local LM Studio model by changing a single environment variable.
LangGraph · Kommo CRM · RAG · LM Studio · Tool calling
A series of end-to-end deployments — regression and classification models packaged and shipped to Google Cloud Run and Azure through GitHub Actions CI/CD, with Streamlit front-ends. Browse the MLOps repos →
PhD in Computer Engineering — UNICAMP · Reinforcement learning for robot skill adaptation
Robot skill adaptation: GMM, SAC-GMM & GMM+PPO on CALVIN — reproducibility study and benchmark of RL methods for skill learning on the CALVIN benchmark, building on Robot Skill Adaptation via Soft Actor-Critic Gaussian Mixture Models (Nematollahi et al., ICRA 2022).
| Method | Type | Status |
|---|---|---|
| GMM only | Offline learning from demonstration (Bayesian GMM) | ✅ Reproduced |
| SAC-GMM | Hybrid — SAC refines trajectory parameters | ✅ Reproduced |
| GMM + continuous PPO | Hybrid — PPO refines Δθ ∈ ℝ⁹ via SB3 | 🧪 In progress |
Evaluated on the open_drawer skill, CALVIN scene D, across 3 seeds × 20 episodes per
method. 📄 Paper in preparation · with Maria Fernanda Paulino Gomes.
📈 Live results, videos & training curves →
- 🔐 Hardening agents for production — layered guardrails, prompt-injection defense and Langfuse-based evaluation loops
- 🧩 Google ADK — exploring Agent Development Kit patterns against LangGraph for multi-agent systems
- 🏠 Local-first LLMs — agents running on LM Studio / Gemma so sensitive data never leaves the machine
- 🦾 PhD research — finishing the continuous GMM+PPO benchmark on CALVIN
Open to collaboration on LLM agents, applied AI and robot learning.