Talk to an engineer

uptime since 2001 · still shipping

Senior engineers for systems that have to keep working.

From 6502 assembly to multi-agent systems.

We build it, we run it, then we automate it. Same team, same client, from the legacy code nobody wants to touch to the agents that now touch everything.

$ claude --mcp drupal An agent trace: asked to create the course "Internal audit 101" with three modules as a draft, the agent calls Drupal four times and reports the draft ready for review.
> Create the course "Internal audit 101" with 3 modules, as draft
⏺ drupal.create_node(type="course", title="Internal audit 101", status="draft")
  ⎿ node/4812 created
⏺ drupal.create_node(type="module", parent=4812, title="Scope and materiality")
  ⎿ node/4813 created
⏺ drupal.create_node(type="module", parent=4812, title="Sampling")
  ⎿ node/4814 created
⏺ drupal.create_node(type="module", parent=4812, title="Reporting")
  ⎿ node/4815 created
✓ Draft ready for review: /admin/content/4812
# a human approves before anything goes live
$ claude --mcp drupal An agent trace: asked to create the course "Internal audit 101" with three modules as a draft, the agent calls Drupal four times and reports the draft ready for review.
> Create the course "Internal audit 101" with 3 modules, as draft
⏺ create_node(course) → 4812
⏺ create_node(module) ×3 → 4813–15
✓ Draft ready for review
worked with


One loop, not a menu.

Drupal, Next.js, DevOps and AI aren't four vendors. They're stages of the same job — so nothing gets lost between "it works on my machine" and "it runs itself".

  1. 01 /build

    Platforms that ship

    Content platforms and web apps on Drupal and Next.js, plus the custom code the problem actually needs.

    • Headless Drupal + Next.js
    • Legacy rescues and upgrades
    • Integrations and APIs

    ideal when: you need it live, not a prototype

  2. 02 /run

    Infra that stays up

    Infrastructure and operations on AWS and Google Cloud. What we build, we operate — and we pick up the pager.

    • Kubernetes, Terraform, Helm
    • CI/CD and release pipelines
    • Monitoring, security, on-call

    ideal when: downtime costs you money

  3. 03 /automate

    Agents on your data

    Agents, chatbots and multi-agent systems wired into the systems you already run — with a human at every gate.

    • MCP servers for your CMS
    • Chatbots over your reports
    • Multi-agent on AWS Bedrock

    ideal when: your team repeats the same task daily


Work

Each case shows which part of the loop we own. Filter by stage.

work/udesa.md

A screenshot of the Universidad de San Andrés work.

Universidad de San Andrés

The university's site, front and back

Headless Drupal and Next.js, built in 2022 and run by us ever since. Next: AI for editors, visitors and marketing.

work/barcelo.md

A screenshot of the Fundación Barceló work.

Fundación Barceló

A new site for a university with campuses across the country

Next.js on a headless Drupal 11, live since October 2026: editors run menus, previews and deploys, with AI suggesting how to tag their content.

$ ls work/ --all · 3 cases →


The AI notebook

No "AI solutions". Just what we've built, in order, like a lab notebook.

Nov 2026

Multi-agent system on AWS Bedrock

AWS Bedrock AgentCore · Step Functions [IN PRODUCTION]
problem:
Business processes that take several people and a lot of manual steps.
what we did:
Coordinated agents on AWS Bedrock.
stack:
AWS Bedrock AgentCore · Step Functions
AWS Bedrockrequestorchestratoragent · 1agent · 2agent · 3result
Oct 2026

AI-DLC as our development workflow

AI-DLC · Claude Code · Kiro [INTERNAL USE]
problem:
Our own harness worked, but it was ours alone: proprietary and hard to standardize.
what we did:
We moved our agent workflow to AI-DLC, an open development lifecycle, to standardize how we work and stop maintaining something proprietary.
stack:
AI-DLC · Claude Code · Kiro
A terminal running the development workflow for the new site: the intent and the rough mockups are done, practices are in progress, requirements come next, and every gate waits for a human.
✓ intent          ✓ rough mockups
▸ practices       · requirements
# every gate waits for a human
Sep 2026

PHP 7 → PHP 8 and Symfony, with agents

PHP 8 · Symfony · Kiro · et-mh [IN PRODUCTION]
problem:
A legacy PHP 7 application on a proprietary framework, with an outdated interface.
what we did:
Ported it to PHP 8, moved it from the proprietary framework to Symfony and built a new look and feel, with Kiro and our own harness (et-mh).
stack:
PHP 8 · Symfony · Kiro · et-mh
A change to composer.json: PHP 7.4 and the legacy framework are removed, PHP 8.2 and the Symfony 7 framework bundle are added.
# composer.json
-  "php": "^7.4",
-  "vendor/legacy-framework": "*"
+  "php": "^8.2",
+  "symfony/framework-bundle": "^7.0"
Nov 2025

Our first development harness (et-mh)

Spec-driven · OpenSpec · coding agents [INTERNAL USE]
problem:
Ship faster without losing control.
what we did:
A spec-driven harness that installs into any repo and works with the coding agent you already use (Claude Code, Codex, Gemini CLI, Cursor): a shared contract, expert agents per stack and human approval gates.
stack:
Spec-driven · OpenSpec · coding agents
A spec written as given, when, then: a visitor on the contact page who submits an empty email sees "Email is required" on the field.
# spec: contact-form.md
GIVEN a visitor on /contact
WHEN  they submit an empty email
THEN  the field shows "Email is required"
Sep 2025

Drupal operated from Claude (MCP)

Drupal · MCP · Claude [IN PRODUCTION]
problem:
Creating courses on the training platform took a lot of manual work.
what we did:
An MCP server on top of Drupal to create and edit trainings from Claude.
stack:
Drupal · MCP · Claude
An agent trace: asked to add a quiz to the course "Internal audit 101", the agent calls Drupal once and links the new quiz to the Sampling module.
> Add a quiz to "Internal audit 101"
⏺ drupal.create_node(type="quiz", parent=4812)
  ⎿ node/4816 created
✓ Linked to module "Sampling"
May 2023

AI chatbot over campaign reports

OpenAI API · Go [IN PRODUCTION]
problem:
Analyzing a campaign's reports meant reading them by hand.
what we did:
A chatbot on the OpenAI API that answers questions about the reports.
stack:
OpenAI API · Go
A question to the chatbot about which ad set had the worst click-through rate last week, answered with set B at 0.4%, citing page 3 of the weekly report.
> Which ad set had the worst CTR last week?
  Set B · 0.4% CTR (weekly report, p.3)
Oct 2020

Anomaly detection with neural networks

Neural network (autoencoder) · Keras · linear regression [POC]
problem:
Find abnormal power consumption among 26,000 records.
what we did:
We trained a neural network (an autoencoder) to learn what normal consumption looks like. Records it can't rebuild well are flagged as anomalies. We compared it with a classic linear fit to see what each method catches.
stack:
Neural network (autoencoder) · Keras · linear regression
inputcompressedrebuilterror high= anomaly
Mar 2020

iArxiv: a machine-learning recommender for physics papers

Machine learning · LDA topic models [PUBLISHED]
problem:
Physicists get hundreds of new arXiv papers every day.
what we did:
A recommender with topic modeling (LDA) that learns each user's preferences. Paper published.
stack:
Machine learning · LDA topic models
Python code that trains a 50-topic LDA model with gensim, to score each day's papers against every reader's topics.
from gensim.models import LdaModel
lda = LdaModel(corpus, num_topics=50, id2word=vocab)
# score today's papers against each reader's topics

We were training autoencoders before it was cool. Now we ship agents.


Three ways to work with us


Got a system that has to keep working?

Tell us what's running, what's breaking, or what you want automated. An engineer — not a salesperson — replies.

mail — hello@easytech.com.ar — 80×24

Talk to an engineer

No sales funnel. The person who answers is the person who'd do the work.

need:

Or write to hello@easytech.com.ar

or write hello@easytech.com.ar
Talk to an engineer