I build AI in regulated domains — eval-first, fail-closed, click-to-verify proof (not slideware).
🇧🇷 Português · 🇬🇧 English · 🇪🇸 Español
| 14 products built, each with a case study | 6 live on their own domains | 3 companies founded and operated from zero |
| 6 OSS modules in NoKey · ~39 automated tests | 59 specialized legal agents in Lastro | 3 languages in this portfolio (PT/EN/ES) |
All verifiable: click the demos below or audit the NoKey repo. No invented revenue numbers — honest rule.
- Recruiter → Numbers above · Evidence map · Bio & thesis
- Client / company → Live proof · Innovation Playbook
- Founder / technical peer → How I build with coding agents · Architecture patterns · Eval methodology · NoKey
| Product | Where | What you verify in 30s |
|---|---|---|
| Assist4Doc ⭐ | assist4doc.com | Record 10s · turns into a structured SOAP · physician must approve or nothing goes to the chart |
| NoKey (OSS) | github.com/kleislm/nokey | Open any demo · DevTools → Network: zero API calls after the model loads |
| Pamella's Hub | pamellamonteiro.com | Portfolio + CMS + LLM lead scoring |
| Staff Forge | pmforge.com.br | Eval suite builder · LLM-as-judge · hands-on AI PM tools |
| Lastro ⭐ | (stealth) | See engineering/architecture-patterns.md — read why a brief with a fabricated case-law citation never ships |
AI innovation hides three things: (1) the manager only runs meetings and can't see ROI evaporate in latency and prompt rework; (2) the engineer only codes and can't see horizon 2/3 or governance; (3) almost no one has the regulatory fluency to put AI in healthcare, law, and finance without creating a liability.
I cover all three — and I prove it on git push.
| Product | The pain it kills 💔 | Horizon | AI stack | Live | Case |
|---|---|---|---|---|---|
| Lastro ⭐ | A legal brief with a fabricated case-law citation = professional liability | H2 — new operating model for legal practice | LangGraph · verified RAG · multi-AI panel · OpenFGA | stealth | 📖 |
| Assist4Doc ⭐ | Physicians lose 3h/day on charts; incomplete notes = clinical-legal risk | H1 at scale — regulated productivity | STT + LLM grounding · human-in-the-loop · multi-tenant | ↗ | 📖 |
| NoKey (OSS) | $$$ on paid APIs that are just compute | H3 — platform optionality | transformers.js · WebGPU · faster-whisper · Playwright · pdf.js | ↗ | 📖 |
| Pamella's Hub | A creator with no CRM, losing leads she didn't even know were leads | H1 — AI in B2C | LLM lead scoring · profile embeddings · headless CMS | ↗ | 📖 |
| Staff Forge | PMs moving into AI PM without hands-on practice | Capability building | LLM-as-judge · curated dataset · regression suite | ↗ | 📖 |
+9 products in the full index (14): TCU Navigator · KLM Legal Hub · ConversaFlow · Insight Health · Agenda Connect · Leone · Desburocratize · OrdemXP · Modulazzi.
Real captures from the live systems, not mockups.
 Assist4Doc ⭐ assist4doc.com · AI clinical documentation |
 Pamella's Hub pamellamonteiro.com · AI CMS + CRM |
 Staff Forge / PM Forge pmforge.com.br · evals + guardrails |
 KLM Legal Hub klmadvogados.com · case management |
 Leone Growth leone.lovable.app · US CRO |
 Desburocratize desburocratize.lovable.app · branding |
An AI feature hides three invisible costs: hallucination that turns into liability, latency that kills UX, and cost-per-token that breaks unit economics. I treat all three as first-class requirements — evals in CI, adversarial verifier in production, task-tier routing, fail-closed when the system can't verify. In a regulated domain, "wrong with confidence" costs the whole company.
- Models — OpenAI · Claude · Gemini · Ollama (gemma3, qwen3, llama3) · WebLLM
- Orchestration — LangGraph · verifier-loop agents · Opus/Sonnet/Haiku routing by cost×quality
- RAG — hybrid BM25 + dense (RRF) · MMR · adversarial verifier · thesis bank
- Backend — FastAPI · APScheduler/arq · PostgreSQL + pgvector · WebSockets · OpenFGA + OPA · Zitadel
- Frontend — React 19 · Vite · TypeScript · Tailwind · shadcn/ui · SSR where it pays for conversion
- Evals & Guardrails — multi-AI panel · LLM-as-judge · regression suite in CI · fail-closed · cost/latency budget
- Infra — Docker · Supabase Edge · blue-green deploy · monthly restore test · OTel tracing
Detail: engineering/ai-stack.md · eval-methodology.md · architecture-patterns.md.
- 🤖 How I build with coding agents — orchestration (execute→review) · CLAUDE.md/skills/sub-agents · token routing · TDD-with-agents · war stories
- 🧭 Innovation Playbook — H1/H2/H3 · governance · responsible AI
- 🧪 Eval methodology — multi-AI panel · LLM-as-judge · regression
- 🏗️ Architecture patterns — verified RAG · verifier loop · fail-closed
- 🗺️ Evidence map — every competency links to the product where it's proven (no self-assigned scores)
- 👤 Bio & thesis — 8 principles · where I want to operate
- 🎓 Certifications — IBM Product + AI Product + Business Analyst trio
Open to conversations about Head of Innovation & AI, Founding AI Engineer, and advisory on AI for regulated domains.
✉️ kleistfilho@gmail.com · 💼 linkedin.com/in/kleist-monteiro · 🐙 github.com/kleislm
Law → growth → founder → ship AI. Brasília, BR MIT-spirit portfolio. Built by Kleist Monteiro.