AI ENGINEER · APPLIED AI · AI SAFETY
Agentic AI · RAG · MCP · Security · Document Intelligence · Data Engineering
LinkedIn · GitHub · AYORAI · Applied Intelligence
I build AI systems where reasoning, authority and execution are explicit engineering boundaries.
Intelligence ≠ Authorization ≠ Execution
My portfolio focuses on Applied AI, AI Engineering and AI Safety, with emphasis on:
Architecture · Security · Evaluation · Reproducibility · Observability · Governance
The work spans agentic systems, RAG, MCP, document intelligence, data pipelines and deterministic controls for AI-assisted execution.
A public-facing intelligence platform combining automated technology ingestion, structured data, AI tooling intelligence and an AI Safety laboratory.
Role in the portfolio: public intelligence and research layer.
AI can produce an answer. ATTRACTOR asks: where is the evidence?
ATTRACTOR is an evidence-verification layer for AI systems that transforms claims into traceable, reproducible and auditable verification results.
AI / Human Claim
│
▼
┌─────────────┐
│ Evidence │
└──────┬──────┘
│
▼
┌─────────────┐
│ Stance │
└──────┬──────┘
│
▼
┌─────────────┐
│ Provenance │
└──────┬──────┘
│
▼
┌───────────────────┐
│ Evidence Clusters │
└─────────┬─────────┘
│
▼
┌───────────────────┐
│ Deterministic │
│ Judge │
└─────────┬─────────┘
│
▼
┌─────────────┐
│ Verdict │
└──────┬──────┘
│
▼
Auditable Record
Core principle: LLM ≠ Judge. The model may help discover or interpret evidence, but final verification is determined by explicit, auditable rules.
ATTRACTOR is designed for scenarios where traceability, integrity, security and auditability matter, with an initial focus on Brazilian telecommunications regulation (Anatel RGC) and LGPD-aware workflows.
It can support governance and validation processes involving evidence, provenance and reproducibility. It does not claim regulatory compliance by itself.
Relevant Brazilian reference: the LGPD (Law No. 13,709/2018). The architecture is also designed to support controls informed by rules issued by financial-system and capital-market regulators, without asserting regulatory compliance by itself.
- Golden v0: 34 cases · 52 documents · SHA-256 manifest (development/evaluation set)
- F1 global-verdict accuracy, always reported side by side (n = 30 each):
| Path | Golden v0 | Golden v0.1 |
|---|---|---|
| A (EVAL_ONLY, research-only) | 40.00% | 76.67% |
| B (commercial candidate) | 43.33% | 66.67% |
| C (rules) | 26.67% | 33.33% |
| Majority baseline | 43.33% | 43.33% |
- On Golden v0, Path B equals the majority baseline. Golden v0.1 was rewritten after per-case results were observed, so it is not independent confirmatory evidence.
- Development sets use synthetic English documents; Portuguese performance is reserved for the hidden Golden v1.
- Judge 2.0 (deterministic adjudication, canonical output, rule trace and generated traceability) shipped in v0.5.0.
- CI: Python 3.11 · 3.12 · 3.13 with Ruff, mypy, pytest, CodeQL and workflow-lint;
mainis protected by required checks. - Release: v0.5.0 · DOI: 10.5281/zenodo.23344005 · all versions: 10.5281/zenodo.23125082
Evaluation classification: A = EVAL_ONLY · B = Commercial Candidate
Do not trust the AI response alone. Verify the evidence.
A security boundary where model intent does not automatically become executable authority.
Agent
↓
Identity / Delegation
↓
MCP Tool Integrity
↓
Runtime Containment
↓
Data-flow Guard
↓
Policy + Authorization
↓
Transaction / Egress Governance
↓
Tool Execution
↓
Provenance / Audit
Focus: least privilege · fail-closed controls · MCP security · threat modeling · adversarial evaluation · auditability.
The three featured projects form a compact applied-AI portfolio:
AYORAI · APPLIED INTELLIGENCE
│
┌─────────────────┼─────────────────┐
│ │ │
Global Tech News AYORAI ATTRACTOR AYORAI AI Shield
Intelligence Evidence-first Runtime defense
│ verification │
└─────────────────┼─────────────────┘
↓
Security + Governance
↓
Authorization / Policy
↓
Execution
↓
Audit / Observability
The recurring architectural boundary is:
Reasoning
≠
Authorization
≠
Execution
| Layer | Evidence |
|---|---|
| Architecture | Explicit components, boundaries and data flow |
| Security | Threat models, least privilege and fail-closed controls |
| Quality | Tests, linting and static analysis |
| Reproducibility | Deterministic examples and documented setup |
| Evaluation | Regression tests and measurable validation |
| Observability | Logs, traces, metrics or audit evidence |
| Delivery | CI/CD and automated quality gates |
| Documentation | Architecture, assumptions, limitations and roadmap |
| Privacy | Public, synthetic or anonymized data |
AI Engineering
Python · LLMs · Agents · RAG · Generative AI · NLP · Vision AI
AI Safety & Security
Policy Engines · Tool Authorization · Least Privilege · MCP Security · Threat Modeling · Adversarial Evaluation · Auditability
Data & Document Intelligence
SQL · pandas · NumPy · ETL/ELT · Data Quality · Power BI · OCR · OpenCV · Tesseract · PDF Processing
Backend & Infrastructure
FastAPI · REST · APIs · JSON · Docker · AWS · Local LLMs · Private AI
Engineering
Git · GitHub Actions · pytest · Ruff · mypy · Bandit · Dependency Auditing · CI/CD · Observability
AYORAI · Applied Intelligence
https://ayorinha.github.io/global-tech-news-ai/
GitHub
https://github.com/Ayorinha
LinkedIn
https://www.linkedin.com/in/anderson-leon-ayora
MBA — Data Science, Analytics & Artificial Intelligence — USP/Esalq
2026–2027
Technology in Database — FIAP
Completed · 2022
This portfolio and its original source code are released under the MIT License.
Copyright © 2026 Anderson Leon Ayora
AYORAI · Applied Intelligence
Applied AI · AI Engineering · AI Safety