Configuration Technician · Portfolio 2026

Configuration Technician · Portfolio 2026

David Gomez

I build AI systems, and applications with them.

From infrastructure, not software

21
MCP tools on one control surface
04
Model routes: Claude, GPT, Gemini, local GPU
08
Projects built with, and on, the fleet

The short version

Configuration Technician by day. The rest of the time I run a multi-node Linux fleet as an agent platform: 21 tools over MCP, jobs routed across Claude, GPT, Gemini and a local model on my own GPU, and nothing marked done until it proves it.

Role
Configuration Technician
Based
Richardson, TX
Status
In service
Fleet
Multi-node
Mesh
Tailscale
Public services
17
Ingress
Cloudflare Tunnel

How I work

Architecture, review, verification

I came to this from infrastructure, not from software engineering. My degree is an AAS in cloud computing, and the programming foundation under it is that degree’s Python coursework plus Harvard’s CS50P. Everything here is built with AI writing most of the code. What I own is the architecture, the machines it runs on, the review of every diff, and the verification: nothing counts as done until it proves it. Knowing what to ask for, spotting when the answer is wrong, and making it survive contact with real hardware is the actual work, and it is why this site is mostly posts about things that broke.

The fleet

See the map →

Control plane
Raspberry Pi
Standby
Raspberry Pi
GPU compute
Dedicated GPU box
Deploy host
Mac mini
Storage
Synology NAS

Selected work

01 to 08 · Proven, not promised · davidisdeploying.com →

01

Tower

An MCP server that turns a multi-node Linux fleet into a pool of AI workers. 21 typed tools, quota-aware routing across three cloud providers and a local open-weight model, and jobs serialised by the resource they touch rather than the machine that runs them.

PythonMCPFastMCPSQLiteAgents
02

Nexus

The operator surface for Tower: live node health, running and finished agent jobs with the evidence each returned, and tiered retrieval over a private corpus. Summary cards first, a separate vector index second, raw sources only for an audit.

FastAPIRAGObservabilityWeb Push
03

Loupe

A self-hosted photo culler with its own ML pipeline: face clustering, scene and aesthetic scoring across roughly 80,000 frames, all running on owned GPU hardware. The demo is the real front end over a fictional library, so it behaves exactly like the logged-in app.

PythonONNXInsightFaceSigLIP2
04

Prospect

A self-hosted job application tracker. A browser extension captures a posting into structured fields as you look at it, normalises inconsistent markup through a per-site adapter registry, and writes into a versioned SQLite schema. Local-first, because that record is not data to hand to a third party.

JavaScriptReactSQLiteExtension
05

Homestead

A house-hunting workspace built on immutable capture. Listing sites edit and pull their own pages, so every capture is archived as evidence first and the comparable-property database rebuilds from it. A delisted property keeps its record.

PythonSQLiteExtensionContent-addressed
06

Waypoint

A study and credential workspace that answers “what should I study next” from recorded progress rather than a static plan, pairing a durable milestone store with a corpus-and-coaching engine behind one authenticated origin.

PythonTypeScriptReactPWARetrieval
08

AWS Capstone

Pricing, deploying and auto-scaling a web app on AWS, then moving its data to a second cloud: the cloud-degree capstone, written up with the numbers.

AWSAuto ScalingMulti-cloud

Latest

All posts →

Aug 19
Making the clean version of a repository the only version that can be pushed
Six of my repositories had gone public straight from their working trees, carrying internal names, addresses, paths and real personal data in test fixtures. Rewriting files does not purge history, so every repo was deleted and recreated from a sanitized single-commit snapshot, and the publishing pipeline was folded into the nightly pusher with a fail-closed gate that is deliberately allowed to tolerate two things.
Aug 18
The agent platform, described in one place for once
I have been building an agent platform on my own hardware one post at a time. This is the whole thing in one place: how a job gets dispatched, which model runs it, and what it has to prove before I believe it.
Aug 17
Twenty hours down, and the watchdog told me why
My control-plane machine vanished for twenty hours. The most useful evidence turned out to be a safety mechanism that did not fire.
Aug 15
Seventy-seven percent of the "edited" photos were not edited
A phone screenshot of an error turned into a day of storage work: an audit of 4,921 pulled photo edits found 3,777 byte-identical to their originals, 108 GB reclaimed, a 3.4 TB library moved with an instant rename, and five separate ways a copy or a delete can lie to you.