Engineering Dossier · Recife, BR → Remote Worldwide

I keep systems alive.
First animals. Now AI.

Fifteen years running a 24/7 veterinary emergency hospital taught me triage, runbooks, and irreversible decisions under pressure. Then I rebuilt my career and applied the same discipline to software — building a full AIOps platform solo and operating 10 production SaaS products on top of it.

AI Reliability Engineer SRE · AIOps · Platform Python · Rust · Django · FastAPI Open to remote contracts
01

The most unusual SRE background you'll read this year

Emergency medicine and site reliability engineering are the same job with different patients. I didn't learn incident response from the Google SRE book — I lived it for 15 years, then found out it had a name.

Emergency OR · 2005–2020
Vital-signs monitoring — HR, SpO₂, blood pressure, baseline deviation
Diagnosis without a talking patient — incomplete information, fast inference
Surgical SOPs — no improvisation under pressure
Triage — critical vs. stable, resource allocation in seconds
Blameless death reviews — what failed in the system, not who failed
Production Systems · 2021–now
Observability — Prometheus baselines, burn-rate alerts, drift detection
Debugging live incidents — partial logs, fast root-cause analysis
Runbooks — executable, tested, versioned
Alert prioritization — P0 vs. noise, escalation policies
Blameless post-mortems — RCA, action items, pattern memory
02

Production vitals

10
SaaS platforms in production
1.17M+
Lines of code, solo-operated
2.938+
Automated tests (AIOps core)
<5min
MTTR — alert to verified fix
100%
Red-team detection rate*
0%
False positives on benchmark*
91%
Test coverage across products
23
Native observability adapters

*DVWA + OWASP Juice Shop benchmark — 16/16 scanner-detectable findings, structural rule: tool output is not evidence.

03

Zenthrus AIOps — the platform I built, then bet my business on

Most engineers use Datadog. I built the closed-loop platform my own products run on: detection, diagnosis, human-approved remediation, verification, and learning — with local-first AI and zero data egress.

# The loop closes in under 5 minutes Alert fired → Aurien diagnoses (RAG + Chain-of-Thought, sources cited) → Human approves (web / Slack Block Kit + HMAC / WhatsApp) → SSH executes runbook (35+ command allowlist, fully audited) → Health check verifies → Alert resolved → Thompson Sampling updates agent reputation (4D reward) → EpisodicMemory stores the pattern → next incident resolves faster
Detection

Predictive, not reactive

Prophet forecasting with multi-signal concept drift (MAPE + KS Test + Page-Hinkley). Multi-window burn rates per the Google SRE Book. Deterministic guards where LLMs don't belong — an absolute architectural rule.

Offense

War Mode red team

Autonomous Kali container (gVisor, cap-drop ALL, egress lockdown, 300s TTL) running 17 offensive tools through an OODA loop, auto-mapped to MITRE ATT&CK. Fail-closed safety kernel with 7 checks.

Resilience

Chaos engineering built-in

SRE Duel: red team injects 15+ failure types, blue-team AI must detect within a 30s SLA. Detection gaps feed back into scan priorities — the defense evolves.

Safety

Digital Twin shadow testing

Every AI-proposed action is simulated on a digital twin before touching production. A Safety Confidence Score gates execution. 21 security layers, 1,471 dedicated tests.

Sovereignty

Local-first AI, air-gap capable

Ollama-served models with pgvector memory and cross-encoder re-ranking. No log, metric, or prompt ever leaves the network — a hard requirement I designed for, not a feature flag.

Systems

eBPF agent in Rust

Kernel-level collection (19 metrics) via a Rust/Aya agent — plus SSH, Prometheus remote_write, and API ingest. Multi-tenant isolation via FK + PostgreSQL RLS + tenant-salted embeddings.

Explore the platform →
04

Not a portfolio. A production fleet.

Every platform below has paying clients and runs under Zenthrus AIOps supervision. Each was commissioned by a real customer, then rebuilt as a scalable multi-tenant SaaS.

AurienConnect
AI-native CRM integrated with ERP ecosystems. Serves a 14-gym network handling ~3,000 leads/month and 30,000 active members.
LIVE · PAYING
ZenthrusBPO
Financial, fiscal and HR BPO platform. In validation with an accounting firm managing 150 corporate entities.
LIVE · VALIDATION
ZenthrusERP
AI-native fiscal & financial ERP built for Brazil's tax reform — NF-e, SPED, intelligent controllership. Rust fiscal engine at its core.
LIVE · PAYING
ZenthrusCar / Locadora / VET
Vertical ERPs for auto dealerships, rental operations, and veterinary clinics — one architecture, tenant-isolated, each with live clients.
LIVE · PAYING
Aurien
Personal AI operations interface: FastAPI daemon orchestrating Claude Code CLI with read-only and TOTP-approved action modes.
INTERNAL · DAILY USE
05

Stack & standards

Core engineering

PythonDjango 5FastAPIRustDjango REST FrameworkCeleryPostgreSQL 15 + pgvectorRedisHTMXReact Native

AI engineering

Claude APIRAG + Cross-Encoder re-rankingOllama (local-first)Thompson Sampling routingLLM guardrails + LLM-as-JudgeProphet forecastingConcept drift detection

Platform & reliability

DockerTraefikPrometheus · Grafana · Loki · TempoeBPF (Rust/Aya)GitHub Actions CI/CDTerraform · GKEMITRE ATT&CKChaos Engineering

Governance as architecture

ISO 27001ISO 42001 (AI governance)SOC 2LGPD / privacy-by-designCompliance-as-Code · 165+ controls
06

How I operate

i

Deterministic guards where LLMs don't belong

Burn rates, thresholds, and safety checks are math, not prompts. AI advises; deterministic code decides what's allowed. This is an absolute rule in everything I ship.

ii

Human-in-the-loop for anything irreversible

Surgery taught me the cost of irreversible actions. Production deploys, remediations above blast-radius thresholds, and credential changes always pass a human gate — MFA-verified.

iii

Evidence over output

My finding classifier enforces a structural rule: tool output is not evidence. It's why the platform benchmarks at zero false positives. I bring the same standard to every claim I make.

iv

Documentation-first, audit-always

ADRs before code. SHA-256 hash-chained audit trails. Five internal audit frameworks run against my own systems every sprint. A scientific background makes rigor a habit, not a chore.

v

AI-augmented, human-accountable

I work daily with Claude CLI as a force multiplier — architecture, review, and delivery velocity of a small team, with one accountable engineer. I'm transparent about it because it's the future of the craft.

Connect

Looking for someone who's calm at 3 a.m.?

I'm open to remote contracts as an AI Reliability Engineer, SRE, or Platform Engineer — async-first teams, worldwide. I've handled emergencies for 15 years. Yours won't scare me.

magno@magnosilva.com LinkedIn Back to site
Recife, Brazil (GMT-3) · Overlaps US Eastern & European hours · English · Portuguese · Spanish (household)