AI Engineer · Data systemsOpen to EU relocation & remoteValencia, VE · UTC−4
Andrés Gutiérrez
I design production AI platforms: agent kernels, MCP integration layers and ML pipelines that business teams actually run on.
6
AI assistants in production on one shared agent kernel
159
MCP tools across 7 servers with per-user OAuth 2.0 PKCE
1,081
unit tests gating every change, plus live-model evals
−95%
alert noise in inventory replenishment
Selected work
Invertropoli C.A. · Oct 2024 – present
Case study · Enterprise AI platform
Sentinel
An enterprise AI platform (FastAPI + React/TypeScript) that puts seven business systems — credit risk, cash flow, spare-parts after-sales, tyre replenishment and orders, fiscal compliance, impact analysis — behind one agent layer, running on Dynamics GP, Odoo and SQL Server for a group distributing GWM and BAIC vehicles.
150→10lines to plug a new system into the agent kernel
7d→livetreasury cash-flow process, now real time
61%of the purchase backlog was invisible to the old KPI
25hMCP crash loop found, root-caused, now watched by health checks
Clients6 assistantsClaude via MCP
↓ role-filtered tools
Agent kerneltool-callingSSEinjection guard
↓
Memory · MLRedispgvectorLightGBMTimesFM
↓ hexagonal adapters
SystemsDynamics GPOdooSQL Server
AI architecture
Shared agent kernel (tool-calling, SSE streaming) reused by 6 assistants: credit, SCM, cash flow, tyres, orders and a federated chatbot.
Integrating a new system dropped from roughly 150 duplicated lines to about 10 lines of config.
MCP layer: 7 servers, 159 verified tools, OAuth 2.0 PKCE so assistants like Claude act on behalf of each user.
The federated chatbot crosses several business systems in a single answer; 116 tools are filtered by role before ever reaching the model — a user without access executes 0 out-of-scope tools.
Agent memory: Redis working memory, pgvector episodic/semantic memory, Jaeger tracing, and per-user token cost control.
6 ML models for credit risk (a risk classifier scoring AUC 0.78–0.81 on a temporal holdout, with SHAP explainability, plus delinquency, days-past-due, credit-limit and amount models), monitored for drift with PSI.
SKU demand forecasting with TimesFM, backtested against ARIMA/STL, feeding an (s,S) replenishment engine.
Business impact
Treasury cash-flow process: from a full week of manual work to real time, with a 40-week rolling projection.
61% of the purchase backlog was invisible to the previous KPI.
Alert noise in inventory replenishment down 95%.
Spare-parts catalogue cleaned by 72%, down to the part numbers actually managed.
Around 7% of the credit portfolio under active early warning.
Tyre order processing: from 8 hours to under 2 (−75%), migrating a fully manual Excel import flow to a native product, via an intermediate Claude + MCP stage.
Real-time fiscal compliance module (tax authority and municipal), removing manual report downloads for a multi-entity group.
Colossus real-time warehouse system: inventory sync from a week to real time, plus tyre picking and forklift routing.
Security & reliability
Found SQL Server connections forcing unencrypted transport in 5 code paths; root-caused to an ODBC driver resolution bug, fixed and verified across all 4 production databases.
Built prompt-injection detection into the shared kernel, protecting all 6 assistants.
Production incident: 7 MCP servers stuck in a roughly 25-hour crash loop unnoticed, caused by an unpinned dependency pulling in a breaking change — root-caused, shipped a minimal safe fix, then added Docker health checks on all 7 services plus a watchdog with real-time chat alerts.
Found 4 of 7 MCP servers not forwarding the real user identity, verified directly against code rather than outdated docs.
CI/CD incident (runner disk full) root-caused to missing Docker cleanup; added a preventive cron job.
Engineering discipline. 1,081+ unit tests plus a separate live-model AI eval suite, hexagonal architecture (domain / application / infrastructure) across 6+ systems, 21 scheduled jobs, RBAC, auditing, Docker and GitLab CI/CD.
AI Analyst — Emerging Technologies Unit
Oct 2024 – present
Invertropoli C.A. — Official distributor of GWM and BAIC vehicles in Venezuela
Design and operate Sentinel end to end: the shared agent kernel, the MCP tool layer, the ML models and the security posture described above.
AI proactive-maintenance system digitizing Excel-based workflows for industrial rotating equipment and vibration reports. Llama 3.2 (1B), served via Hugging Face, automates FMEA and maintenance recommendations. Built an automated scheduling module integrated with the chatbot (PHP, MySQL, JavaScript). Also the subject of his degree thesis.
Open source
github.com/ElSirGuti
Data pipeline · ML
Motorsport Analytics Pipeline
Upload CSV lap data to compare driver performance lap by lap, analyze time deltas, generate G-G diagrams, model tyre and brake temperatures, flag anomalies with an Isolation Forest, and get setup advice from an AI advisor.
Development of an AI-based proactive maintenance system for RSC Services C.A.: Llama 3.2 reads vibration reports, automates FMEA and schedules maintenance for industrial rotating equipment.