Medical appointment booking platform
From zero to production for a private client: Next.js 14 with App Router, Prisma on MySQL, Zustand state and in-house auth. Still running today.
Next.js 14 · Prisma · MySQL · Zustand
Fig. 1 · Platform Engineer · Full Stack · Mexicali, MX
Self-host
every-
thing
I build the product, set up the pipeline that ships it and run the hardware it lives on. This site is all three: it runs on the Raspberry Pi marked in the blueprint, at my home.
Blueprint no. 001 · homelabonline
UPTIME
—
CPU TEMP
—
LOAD 1m
—
TTFB
—
Fig. 2 · The argument
Most companies pay for cloud they do not need.
The cloud makes sense when load is unpredictable or when nobody is around to run servers. For a steady workload, a dedicated box or a well-run rack costs a fraction, and sensitive data never leaves the building. I have done it with ERPs, with AI servers, and with this page.
86 %
of CIOs plan to move workloads from public cloud back to their own infrastructure. An all-time high.
Barclays CIO Survey, 2025
$7 M
saved over five years by 37signals after leaving AWS. They spent $600 K on servers.
37signals, 2024
59 %
of organizations spent more than budgeted on cloud in 2024.
Flexera State of the Cloud, 2025
$0/mo
hosting cost of this page. A Pi I already owned, a free tunnel and under 10 W of power.
Measured in my homelab
Fig. 3 · Own projects
Two systems that run on their own, 24/7
Project 01 · fintech / quantitativein production
Trading AI
A quant trading bot that learned not to use AI.
24/7 paper trading on Binance, five strategies in parallel and a home-built backtest engine that discarded 240 variants, the LLM version included, before letting the winner run. Runs on this very Pi.
- frontend ✓
- backend ✓
- data ✓
- ops ✓
FK19 $60 → $90LLM $60 → $56shaded: out-of-sample
Win rate OOS
76 %
Profit factor
2.49
Max DD OOS
2.6 %
AI cost
$0
node · binance ws · sse · docker arm64$ cat trading-ai.md →
Project 02 · infrastructure / observabilityin operation
Control room
Per-process routing, automatic failover and frozen-instance detection.
32 instances of the same executable on one Windows box, each on its own network egress, without injecting code into the application. Deploys itself from git and rolls back when something fails.
- frontend ✓
- backend ✓
- data ✗
- ops ✓
1/7
0
0
0
0
Instances
32
Buckets
7
Deploy
≤10 min
Injected code
0
powershell 5.1 · p/invoke · kernel routing · git deploy$ cat control-room.md →
Fig. 4 · In production for others · 300-user company, 2023 to 2026
What I shipped in production, by layer
01 / Buildfrontend · backend · data
High-traffic Laravel platform, twice as fast
Query refactor and strategic caching that halved response times. Moved the critical sections from Vue 2 to Alpine.js with a smaller bundle.
Laravel · Eloquent · Alpine.js · Redis
02 / Automateci/cd · containers · workflows
Self-hosted GitLab with CI/CD for the team
In-house repository with pipelines for development. Dockerised the existing Node and Laravel apps: identical deploys everywhere and less time to onboard a new service.
GitLab CI · Docker Compose · pnpm workspaces
Leads classified by a local LLM, no manual work
n8n workflows wired to Ollama on the company AI server. The sales team stopped triaging by hand and no data ever left for an external provider.
n8n · Ollama · internal APIs
03 / Operatevirtualisation · network · hardware
From the cloud to Proxmox, 300 users
Full migration to a dedicated on-premise server: 5 VMs, 1 LXC and dozens of containers. ERP with 200 concurrent users, scheduled backups and RAID10.
Proxmox VE · ERPNext · RAID10
A 6-GPU AI server and a network that stays up
Physically built the 6× RTX 3090 server for local inference. Network for 300 users with 5-ISP multi-WAN failover, VLANs per critical area and WireGuard VPN.
6× RTX 3090 · OPNsense · WireGuard