DEX
Multi-agent orchestration · On-prem / sovereign Ancona · Marche · Italy

WITHOUT
THE CLOUD

I build and run multi-agent systems that work without the cloud — verified deploys, on hardware you own.

Alias

deucelenon / Deuce Lenon

01

The position

Self-hosted agent orchestration that runs on hardware you already own — verified before it ships.

Constraint, not scope

This is not a claim to do everything. It is a claim under a constraint: a full multi-agent runtime with no cloud dependency, running on constrained local hardware.

Why it matters

Some data cannot leave the building. Banks, healthcare, government and anyone with a residency or sovereignty requirement need agents that run locally — not on someone else’s servers.

How it is proven

A working system ranks above a claim — a live URL, a public repo, reproducible metrics, and a disclosed method. Claims are kept at or below what the artifact proves.

Based
Ancona, Marche, Italy
Role
Website Designer & Custom App Developer
Focus
Agentic AI · multi-agent orchestration
Studies
Digital Economics & Business, UnivPM (2023–2026)
02

The sovereign constraint

Data cannot leave. So the agents run where the data already is.

On-prem by design

The runtime is isolated and self-hosted. No cloud dependency, no data sent to a third-party inference service — the system lives inside your perimeter.

Hardware you own

Deployed to the machine you control: an owned server, a VPS, or constrained local hardware. You keep the box, the data and the keys.

Verified before it ships

Deploys are verification-gated: a change ships only once it passes the checks. The pattern is describable and repeatable, not a slogan.

03

Productised service

Priced as an outcome, delivered as a managed service — not billed by the hour.

Tier 01

Deploy

A self-hosted agent system installed on your hardware, verified before handover.

  • On-prem installation on the machine you own
  • Multi-agent runtime wired to your goal
  • Verification-gated deploy before handover
  • Reproducible load test with disclosed method
Scoped per deployment
Tier 02 · Managed

Run

The system, operated and kept running. This is the default engagement.

  • Everything in Deploy
  • Ongoing operation and monitoring
  • Updates and changes shipped through verification
  • Direct line for incidents and requests
Monthly managed service
Tier 03

Extend

Additional agents, integrations and workflows added to a running deployment.

  • New role-specialised agents
  • Integrations to services you already use
  • New workflows on the same runtime
  • Same verification gate as day one
Scoped per engagement

Why outcome-priced, not hourly. The value is a working system in production, not hours spent. Hourly rates in this market compress toward the bottom — multi-agent orchestration is crowded there. A productised managed service is priced against the outcome and the constraint it removes, and it is the model the on-prem / sovereign buyer is actually buying.

04

Verified work

Gestione Plaza

plazaancona.it dex.plazaancona.it
Menu

Bilingual live menu

132-item Italian/English menu with live price editing, photo upload and crop, and section reordering.

Reservations

Floor-plan bookings

Online reservations with a 3-zone floor plan — upstairs, indoor, outdoor — and table availability.

Orders

Live order pipeline

Atomic order numbering with a two-way guest-to-staff chat, tracked by conversation numbers.

Alerts

Instant staff alerts

Every order, booking and message pushed to staff phones via ntfy.

Context

Real production system

A production website plus staff admin for a real restaurant and cocktail bar.

Stack

Lean and self-hosted

Python (stdlib http.server), vanilla JS/HTML/CSS, Caddy, systemd, Cloudflare, ntfy.

200/200Requests OK under load
0Errors
~60 rpsOn a 1-core / 1.8 GB VPS

Verified under load: 200 concurrent requests returned 200/200 OK with 0 errors, at roughly 60 requests per second on a 1-core / 1.8 GB VPS — the method disclosed, the number reproducible.

JIT Framework v2.0

klawskill.com
Input

Plain-language goal

A goal written in everyday language is the entry point to the system.

Pipeline

Analyse · select · execute

Task analysis, then skill selection, then execution — just in time.

Status

Published and working

A working published system — not a demo.

05

Honest limits

Stated before you find them — because the limits are part of the offer.

01

Single host

One physical device: no redundancy, no failover. That is a deliberate choice for sovereignty, and it is a hard availability ceiling. If you need multi-region redundancy, I will tell you to go elsewhere.

02

Serialised writes

Concurrent agent runs are serialised on a shared SQLite write-lock. The honest upper bound is a measured number on this hardware, not a guess — so it gets measured, and stated as measured.

03

One case study

Gestione Plaza is the one real, verifiable client case study. That is exactly what caps how enterprise a claim can be — and why the work is priced productised and outcome-based, not as an enterprise retainer.

06

Focus

Where the work is pointed.

  • 01Self-hosted multi-agent orchestration
  • 02On-prem / sovereign agent deploys
  • 03Verification-gated releases
  • 04Websites & custom applications