Hi, I'm Henrico Piubello 👋
AI Engineer. I build end-to-end AI products, from data to deployment, and lead engineering teams when needed. 📍 Juiz de Fora, MG, Brazil · 🌐 hepine.com · ✉️ contato@hepine.com
From industrial maintenance to LLMs in production. Time and team size are no longer the limit. The only limit is the size of the idea.
- 🧪 Head of the AI Lab at Núcleo de Performance — leading the technical side of client work across offline conversions, attribution, websites, e-commerce migrations, and incident response
- 🛠️ Technical Lead at Starhash — a live platform that licenses creators' likenesses for generative AI
- 🤝 Partner, AI Engineer, and Tech Lead at Solabs — building SolabsForce, an omnichannel CRM and customer service platform powered by AI agents
- 📚 Previously: 5.5 years at Voitto, progressing from intern to senior specialist and leading the development, sales, and IT teams
| Project | What it does |
|---|---|
| Realidade Brasil | Civic tech for Brazil's 2026 elections: a blind-voting experience built from real excerpts of proposals registered with the electoral authority. Shipped from zero to production in about six days, with 192 tests and 50 end-to-end scenarios. |
| Cumpre | AI that reads contracts, checks interest rates and deadlines, and sends reminders before due dates — while only making claims it can support with exact quotes. |
| CodeCrush | A Brazilian developer hub I have maintained since 2023, featuring a blog, 40 tools, a job board, an AI radar, and career maps. |
| WifiShow | A native macOS Wi-Fi spectrum analyzer built with SwiftUI. See who occupies each channel, at what width, and with what signal strength — open source, dependency-free, and without network access. |
- Unhead — escaped control characters in JSON template parameters
- AnythingLLM — moved Azure OpenAI reasoning-agent calls to the Responses API
- Keep — added alert labels to the alert sidebar
- Hermes WebUI — preserved provider identity in fallback-model badges
- DreamCoder — constrained chat image attachments
- Agent Swarm — documented model-tier selection for task execution
These pull requests are open for review; they have not been merged.
| Project | Contribution |
|---|---|
| LlamaIndex | Restore temporary attributes when a setter fails. |
| Langflow | Enter the asyncio timeout context asynchronously. |
| Pipecat | Count supplementary CJK ideographs as spoken words. |
| Braintrust | Await PromiseLike results before ending traced spans. |
| React Hook Form | Clear pending validation when a resolver rejects. |
| Redux Toolkit | Append query parameters before URL fragments. |
| Svelte | Recognize case-insensitive CSS keyframes at-rules. |
| Material UI | Preserve independent sx with styled-components. |
Full contribution record, including merged PRs, open reviews and drafts. Status verified on November 7, 2026.
I also publish reusable automation in henricop/skills, including GitHub workflow, visual prototyping, and local transcription skills.
- Multiple agents, one contract — Claude Code, Codex, and OpenCode work in the same repositories and follow the same
AGENTS.md - Parallel execution and comparison — agents work on the same problem concurrently, outputs are compared, and responsibilities are divided by task type
- Review and guardrails — rigorous review and testing; one hook forces a workflow backup before any n8n change
- Custom tools — skills I built for daily handoffs that reconcile local AI activity with Jira and Asana, delivery journaling, and AI-ready written brand governance
