Cortex Resolve — Agentic Conversational Platform
In Production
What
Agentic AI platform handling live customer operations for a major Brazilian public utility serving 1.5 million people — deployed inside the client’s business as their public-facing assistant, not a lab experiment.
Architecture
The LLM reasons over structured retrieval and emits authenticated JSON-RPC tool calls through an MCP layer (about 50 semantic tools). Client systems execute the calls and remain the systems of record; human escalation is retained.
Evaluation
Approximately 8,000 interactions in the first two months, with 99%+ successful resolution. These figures describe that observation period, not the utility’s active use of the platform.
My role
Founder & CEO of the company that built it: product direction, client relationship, commercial terms, and outcome accountability.
Windows Compliance Agent Suite
Shipped at Microsoft
What
Production agents (Reg2Req · Regulation Intake · PoliCheck) that transformed regulatory compliance work across 200+ Windows engineering teams: obligations ingested and translated into sequenced engineering requirements, policy checks automated, human-in-the-loop review grounded in source documents.
Architecture
LLM + RAG over regulatory and policy corpora; prompt orchestration; human-in-the-loop validation with corporate-legal-validated evidence records; Responsible AI governance throughout.
40–60% program cycle-time reduction
30–40% manual validation effort recovered
Zero major audit findings, shipping monthly
Global Readiness: failure, evaluation, rebuild
Microsoft · 2025–2026Evaluation
The prompt-based v1 falsely approved imagery; downstream legal review exposed the failure. I re-reviewed manually, rebuilt the agent in Copilot Studio on retrieval grounded in approved images and policy standards, and tested against known failures before return to service.
Outcome
Readiness review fell from 7–9 hours to under 40 minutes, retaining human final approval. Reg2Req separately scanned trusted regulatory sources and routed about 12 vetted applicability-assessed updates per day; its 40%+ cost reduction is an estimate.
BCL Agent — Continuity Review Automation
Shipped at Microsoft
What
Automated business-continuity review agent replacing manual assessment toil across hundreds of services, with human review retained where judgment matters. Adoption was earned by design — trust built in, not mandated.
Up to 90% of review workload automated
~83% measured time savings per assessment
WorkMind — Clinical Assessment Platform
Pilot completed
What
Occupational mental-health assessment platform for regulatory compliance (Brazil NR-1): validated clinical instruments, psychiatrist sign-off authority, consent architecture, and k≥5 anonymity thresholds engineered in (LGPD).
Why it matters
AI-adjacent product where governance is existential — evidence that I build for stakes, not demos. Companion work: co-author of A Nova Era da Saúde Mental com Inteligência Artificial (Artmed, 2026), with Dr. Antônio Geraldo da Silva and Dra. Ana Luiza Silva Teles.
Daily Practice
Ongoing
How I work
Agents, RAG pipelines, and prompt orchestration are my default workflow for analysis, drafting, and build acceleration — across enterprise and founder product work. Platforms: Claude, Copilot Studio, Azure AI Foundry, Azure OpenAI, Power Automate.