All Projects

Overnight AI Build Lab

I close my laptop. The build keeps running.

Private build environment.

An amber line passes through checkpoints and a server to a build report on a desk at night.

Telegram → OpenCode

Send a brief. Wake to a report.

The build runs on a remote VM.

Saved checkpoints let it pick up where it stopped.

A Telegram brief starts a checkpointed OpenCode build on a remote VM. The bot sends back a report for review.

Tools and Skills

Technologies

  • Telegram
  • Podman
  • OpenCode
  • Tailscale
  • Ubuntu

Skills Applied

  • Agent orchestration
  • Checkpoint recovery
  • Rootless containers
  • CI/CD and rollback

The problem

I wanted to test small product ideas without spending every evening starting from an empty repo. The build needed to run independently of my laptop, preserve progress, and leave me enough evidence to review the result in the morning.

What I built

I set up one Ubuntu ARM64 VM on Oracle Cloud for remote development and overnight builds. During the day I connect over Tailscale with VS Code Remote SSH. For an overnight build, I send a brief to a Telegram bot. OpenCode agents plan the work, implement it, run checks, and review it before sending a report. Systemd keeps the bot and preview services running.

Why I built it this way

Use Telegram for input and progress.

I can submit a brief or check a run from my phone. Code, logs, and artifacts stay on the VM.

Save checkpoints between tasks.

The plan and task ledger show what finished and where the run stopped. A recoverable failure can retry or switch model providers. A blocked task can report what happened and wait for me to intervene.

Give builds and previews their own runtimes.

Builds run inside rootless Podman. Published previews use a separate runtime with CPU and memory limits. Provider and signing credentials stay on the host.

A build from the lab

The Supplier Invoice Reviewer checks invoices against purchase orders and contract terms. In the demo, both the invoice and purchase order listed $12 per unit. The contract listed $10. Across 100 units, the reviewer flagged a $200 discrepancy and pointed to the invoice row and pricing amendment.

Prototype previews can use Tailscale Funnel. The invoice reviewer has a separate path: GitHub Actions tests and builds the React app, sends the artifact to the VM over restricted SSH, and switches the Caddy-served HTTPS release. It retains the prior build for rollback.

Have a workflow you’d like to prototype?

Bring one task that can be defined and checked independently. We’ll work out what a first run should produce and how you’ll know it succeeded.

Book a Discovery Call