Skip to content
ANDRES QUINTERO
Leer esta página en español
Portrait of Andrés

ANDRES QUINTERO · Valencia, Spain

AI Engineer · Automation and self-hosted infrastructure

22 years hands-on with systems, since I was 14 · Open to employment or projects · Valencia, hybrid or remote

From the agentthat answers to theserver it runs on.

Let's talkView my CV

I build conversational agents, automate processes and run the infrastructure they live on. When the data is sensitive, all processing happens inside the client's perimeter: local models, a segmented network and not one byte to third parties.

What I do

Four layers, one person.

Most people cover one. The value is in having the conversation, the process, the data and the server designed by whoever will answer when something breaks.

01Conversation

AI agents and chatbots

Conversational agents over WhatsApp, web and phone, wired into the client's own APIs and CRM, with RAG and purpose-built calculation functions.

02Processes

Process automation

Self-hosted n8n, with close to twenty flows in production, integrating Odoo, Holded, Pipedrive, payment gateways and telephony.

03Product and data

Product and development

Complete applications end to end: multi-tenant SaaS, a FastAPI backend with hashed keys, signed URLs, sessions and an audit trail, and the interface on top.

04Server

Self-hosted infrastructure

Proxmox, OPNsense and Docker on our own servers: VLAN-segmented networks, WireGuard VPN, HAProxy publishing, hardening and audit trails. No SaaS dependency and no third-party processors to add to your GDPR record, because the data never leaves your perimeter.

Projects

Systems in operation.

Work done for third parties is described by sector, without naming companies.

13 projects shown

My own projects

06
In productionInternal infrastructure · 2021 – present

Private AI infrastructure

Using AI on sensitive data normally means shipping it to an outside provider. I wanted the opposite: full AI capability without a single record leaving the perimeter.

An owned inference and AI services server, built and operated end to end. Virtualisation with a dedicated GPU in passthrough, and language, voice and document-search models running locally behind a model proxy. Network segmented into VLANs behind my own firewall, remote access over an encrypted tunnel and services published through a reverse proxy. It is the machine everything else I build runs on, and the proof the approach holds: the same result as a cloud service, without the data leaving.

Infraestructura de IA privada

Infraestructura de IA privada Una consulta con dato sensible entra por la red segmentada, se resuelve contra los modelos y el índice documental que corren en la máquina propia, y la respuesta vuelve al que preguntó sin que el dato haya salido del perímetro. QUÉ PIDE EL NEGOCIO Consulta sensibleTexto, voz o documento 01Red segmentadaSin salida al exterior 02Inferencia en localLenguaje, voz, documentos 03Búsqueda documentalÍndice construido en casa 04Respuesta compuestaCon su fuente Vuelve a quien preguntóSin salir del perímetro El dato no sale del perímetro del cliente
  • Proxmox
  • GPU passthrough
  • Local models
  • OPNsense
  • WireGuard
  • HAProxy
In productionOwn subscription business · 2017 – present

Recurring billing and payment validation

Billing a monthly subscription to hundreds of customers meant chasing invoices one by one and hand-checking every payment receipt that arrived over messaging.

A daily automatic sweep of active unpaid invoices, with a personalised reminder to each customer over WhatsApp and email, with no human in the loop. Receipts sent in by customers are processed with OCR: the destination account is checked, approval number and date are extracted, and the payment is recorded, or the exception is logged with its reason. Any failure in the process raises an alert immediately.

Cobro recurrente y validación de pagos

Cobro recurrente y validación de pagos Dos recorridos diarios: el aviso de las facturas pendientes por mensajería y correo, y la lectura de los comprobantes que envía el cliente hasta quedar registrados o marcados como incidencia con su motivo. POR DÓNDE ENTRA Llega el comprobanteFoto o captura 01Lectura del comprobanteCualquier formato 02VerificaciónCuenta, aprobación, fecha RegistradoConciliado IncidenciaCon su motivo Nadie persigue un cobro ni comprueba un recibo
  • n8n
  • OCR
  • LLM
  • MySQL
  • InvoiceNinja
Wound downOwn service · Jan 2016 – Aug 2017

Subscription proxy service

Running a paid service for more than 150 customers, billed by connected device count, while competitors aimed denial-of-service attacks at the infrastructure.

Infrastructure deployed and operated entirely by me on AWS. Service continuity under attack: address rotation, migration to other servers and recovery without cutting customers off. Sign-ups, cancellations, support and recurring billing all handled end to end, single-handed.

Servicio de proxy por suscripción

Servicio de proxy por suscripción El alta por suscripción reparte accesos según los dispositivos contratados y el servicio corre sobre infraestructura propia bajo vigilancia continua. Cuando llega un ataque de denegación, la rotación de direcciones y la migración de servidores mantienen el servicio en pie sin cortar a los clientes. POR DÓNDE ENTRA Alta por suscripciónCobro por dispositivos 01Reparto de accesosSegún lo contratado 02Servicio en marchaInfraestructura propia 03Llega un ataqueLa vigilancia lo detecta 04Rotación y migraciónDirecciones y servidor No pierde servicioNi se entera del cambio La continuidad se sostiene sin cortar a nadie
  • AWS
  • SOCKS5
  • Linux
  • Server administration
In productionInternal infrastructure · 2026

Infrastructure operations agents

Keeping an infrastructure's state and security current demands constant review that, in practice, nobody performs as often as they should.

A team of agents running local models inside an isolated virtual machine, working against the hypervisor and firewall APIs. They do more than audit: they detect changes and security findings, create virtual machines, apply rules and deploy services. No action that writes runs without prior human approval, and every one lands in a local Git repository that works as an auditable record of what was done and why.

Agentes de operación de infraestructura

Agentes de operación de infraestructura Una revisión programada o un aviso despiertan al agente, que lee el estado de la infraestructura desde una máquina aislada y propone el cambio. Una persona lo aprueba antes de que se aplique, y toda la actuación queda registrada. QUÉ LO DISPARA Revisión programadaO un aviso 01Lee el estadoMáquina aislada 02Propone el cambioQué tocar y por qué 03Una persona apruebaNada pasa sin esto Se aplicaReglas y servicios Queda en el historialQué y por qué Nada se aplica sin aprobación ni sin rastro
  • Proxmox
  • OPNsense
  • Local models
  • Human approval
  • Git
In developmentTelephony · 2026

Self-hosted voice agent

Commercial voice-agent platforms charge by the minute and leave recordings and transcripts in a third party's hands.

Transcription, language model and speech synthesis all running locally, on a self-hosted PBX with a SIP line. The same capability as the commercial platforms, with no per-minute cost, no vendor dependency, and recordings kept inside the perimeter.

Agente de voz telefónico autoalojado

Agente de voz telefónico autoalojado La llamada entra por una centralita propia con línea SIP. Transcripción, modelo y síntesis de voz se ejecutan en la misma máquina, así que la conversación se resuelve sin coste por minuto y sin que el audio salga hacia un tercero. POR DÓNDE ENTRA Llamada a la líneaCentralita propia 01Se transcribe al vueloEn la propia máquina 02El modelo respondeCon el contexto 03Se sintetiza la vozY vuelve a la línea Conversación resueltaSin coste por minuto Nada sale fueraNi audio ni texto Ni coste por minuto ni audio en manos de terceros
  • SIP
  • Self-hosted PBX
  • STT
  • Local LLM
  • TTS
DeliveredReal estate, own engagement · Feb – Mar 2025

Phone agent for lead qualification

The volume of information calls was eating the sales team's time.

A phone agent backed by the property database, with RAG over the conversation script. It works out which property the caller means from the address or the listing number. It qualifies, then hands the case to a second agent with its own RAG of criteria and a function that computes viability from income. The answer comes back within the same call.

Agente telefónico de cualificación de leads

Agente telefónico de cualificación de leads La llamada entra, el agente identifica el inmueble contra la base de propiedades y conversa para recoger los datos. Un segundo agente calcula la viabilidad según los ingresos y la llamada termina con el interesado cualificado o descartado con su motivo. POR DÓNDE ENTRA Llamada de un interesadoPortales y redes 01Identifica el inmuebleContra la base 02Conversa y recoge datosGuion asistido 03Calcula viabilidadSegún ingresos Apto, con su fichaListo para visita No aptoCon el motivo La cualificación termina en la misma llamada
  • Vapi
  • RAG
  • Property database
  • Calculation functions

Work done as an employee

07

  • Automated sign-up and billing

    DeliveredCooperatives · Oct 2025Featured

    A single chat takes the lead from the first question to full membership. It answers with RAG over the terms, filters for eligibility and, if they qualify, offers to continue: the Stripe payment widget appears inside the conversation itself. Once paid, it collects data and documents, runs OCR on them and catches the problems before a person would — an expired ID, a field that does not match — so the member can fix it on the spot. If everything checks out it registers them on the cooperative's platform; if not, it leaves them pending with a note on what is missing.

    • n8n
    • RAG
    • Stripe
    • OCR
    • Pipedrive

  • Multimodal customer support over WhatsApp

    DeliveredIndustrial distribution · May 2025

    WhatsApp chatbot that replies with synthesised voice, images and video, and also listens to and transcribes the audio the customer sends. It generates quotes by integrating directly with Holded.

    • n8n
    • ElevenLabs
    • Audio transcription
    • Holded

  • Energy bill comparison

    DeliveredEnergy · Aug 2025

    A chatbot that takes the customer's latest bill, reads it, and compares it against the retailer's API to establish whether a better price can be offered.

    • Document extraction
    • Tariff API
    • n8n

  • On-premise server and infrastructure

    DeliveredInternal infrastructure · Nov 2025

    Build-out of the company's physical server on Proxmox and OPNsense, with segmented networks and secure remote access over VPN. Self-hosting of n8n and the rest of the stack, with services published by subdomain, leaving the infrastructure ready and free of recurring SaaS cost.

    • Proxmox
    • OPNsense
    • Docker
    • WireGuard
Tools

What I work with.

Languages and data

  • Python
  • TypeScript
  • SQL
  • PostgreSQL
  • MySQL
  • Redis

AI

  • Agents
  • RAG and semantic search
  • Local models (Ollama, vLLM)
  • Vertex AI
  • OpenAI
  • Anthropic
  • LiteLLM
  • STT and TTS
  • OCR
  • MCP

Automation and integration

  • Self-hosted n8n
  • Selenium
  • Playwright
  • REST APIs and webhooks
  • Odoo
  • Holded
  • Pipedrive
  • Payment gateways

Infrastructure

  • Proxmox
  • Docker
  • Linux
  • OPNsense
  • OpenWRT
  • WireGuard
  • HAProxy
  • AWS
  • Google Cloud
  • Hetzner
  • Coolify
  • ZFS

Security

  • Network segmentation
  • VPN
  • Hardening
  • Secrets management
  • Anonymisation (Presidio)
  • Audit trails
Background

Before this.

  1. The first laptop

    I started at fourteen, when I was given my first laptop. Then came computing books of every kind, and the habit of opening things up to see how they are built inside.

  2. The automating habit

    Everything else came from there. Years before AI was a topic, I was already building reminders and automatic alerts with Python and Selenium, driving WhatsApp Web remotely because there was no other way: repetitive work has always looked to me like a technical problem, not an obligation.

  3. My own businesses

    Then came my own. A restaurant that grew to seven employees, run on Odoo I self-hosted, with the cameras installed by me and customer acquisition automated on social media. And a subscription service with hundreds of paying customers, on my own AWS infrastructure.

I have learned all of it by building things I needed.

What I now do for others is exactly that.

Portrait of Andrés
Contact

Shall we build it?

If you have a process that repeats every day, or a system you want to bring onto your own infrastructure, write to me and we will take a look.

Compliant without losing the conversation

Since August 2026, Article 50 of the EU AI Act requires that whoever calls or writes knows they are being served by an automated system. It sounds harsher than it is: one well-written opening line settles it and the conversation carries on, with nobody hanging up. I design it that way from the start. And the other half comes for free: because the processing happens inside your own infrastructure, there are no transfers to justify and no third-party processors to add to your record.

andrestech77@gmail.com

Valencia, Spain