IT Infrastructure & Automation Engineer — Buenos Aires, Argentina 🇦🇷
Since July 2026 I have been on the IT team of a six-company automotive dealer group: 30+ branch offices, 500+ Windows endpoints, six independent Active Directory domains. Beyond operating that environment, I build tooling for it.
Most of my day is Windows, Active Directory, networking and PowerShell. The interesting part is what sits on top: an in-house RMM, internal MCP servers over PostgreSQL and SQL Server, and an LLM agent gateway where every security-critical path runs as deterministic code instead of model output.
In-house RMM — PowerShell agent running as SYSTEM on every workstation, PostgreSQL back end, Node/Express API, React dashboard. Hardware and software inventory, antivirus and disk alerts, online/offline state, remote command execution. 500+ endpoints across six companies.
Agent platform — an operations gateway on an LLM agent runtime: one orchestrator delegating to four isolated specialist agents, each with its own workspace and an explicit tool deny-list. Structured queries moved from 60–150 s of model inference to under 2 s by intercepting intent in code before the model ever sees the message.
MCP servers — five of them, exposing PostgreSQL and SQL Server to agents through fixed parameterized queries only. No free-form SQL reachable by the model, password columns excluded by construction, table allowlist validated in code rather than stated in a prompt. The pattern, rebuilt from scratch with its tests, is public: guarded-sql-mcp.
Provisioning tooling — Active Directory and Google Workspace account operations (Express, ldapjs, Google Admin SDK), with confirmation required before anything writes. A PowerShell take on the AD side, written from scratch for the public repo, is here: ad-lifecycle.
| Project | What it is | Stack |
|---|---|---|
| SystemMonitor | RMM console and agents: inventory, metrics, remote execution with every command audited against its operator | React · TypeScript · Node/Express · PostgreSQL · PowerShell |
| guarded-sql-mcp | MCP server exposing PostgreSQL to LLM agents through a fixed catalog of read-only, parameterized queries | TypeScript · MCP SDK · PostgreSQL |
| ad-lifecycle | PowerShell module for AD joiner / mover / leaver operations, -WhatIf on every write, Pester-tested |
PowerShell · Pester |
| AssetDesk | IT service desk and asset inventory with server-enforced roles and a ticket lifecycle · live demo | React · Vite · Node/Express · SQLite |
| CommerceSuite | Procurement storefront: server-side totals and stock, order state machine, role-gated admin · live demo | React · Vite · Node/Express · SQLite |
| Portfolio | Personal site · live | React · Vite · TailwindCSS |
Infrastructure Windows Server · Active Directory (LDAP, OUs, GPO) · DNS · DHCP · TCP/IP · multi-site VPN · Linux (Ubuntu, Debian)
Automation PowerShell · Python · Bash · Node.js · Windows Task Scheduler
Development TypeScript · Node.js/Express · React · Next.js · TailwindCSS · REST APIs · PostgreSQL · SQL Server · SQLite · Git
AI engineering Multi-agent orchestration · per-agent tool policies · MCP server development · deterministic pre-inference interception · Ollama · Gemini API
Operations PDQ Inventory · Kaspersky Endpoint · Zammad · OTRS · Google Workspace Admin SDK and GAM
Worked with, not claiming depth: Docker · MikroTik · FortiGate · Hyper-V · Prometheus · Grafana · OpenVPN
Read-only by default. Explicit confirmation before anything destructive. Access rules enforced in code, not written in a prompt and hoped for. Every system documented well enough that the next person doesn't have to rediscover it — including the parts that are broken and why.
Remote roles in IT automation and infrastructure engineering, applied AI / agent engineering, or platform and internal-tooling work. Based in UTC−3, overlapping a full working day with US Eastern and Central hours.