Case Study

24/7 AI Crypto Trading Livestream

Unattended multi-platform video pipeline broadcasting real-money trading bots around the clock, with trade-triggered auto-clipping and self-healing watchdogs.

Live in productionVideo Pipelineffmpeg · Chromium · Xvfb24/7 uptime

The brief

My own project: I run real trading bots on real capital (Kraken + Alpaca) and wanted to broadcast that activity live, continuously, as a build-in-public channel — not a pre-recorded highlight reel. That meant a video pipeline with no human operator: no one cutting scenes, restarting a stuck encoder, or babysitting a crashed browser tab at 3am.

Constraint: run on hardware already in the house (no cloud encoding budget), stream to multiple platforms from one source, and survive indefinitely — days and weeks of continuous uptime — without a person watching it.

The approach

The "camera" is a headless Chromium browser under Xvfb rendering an HTML/JS overlay dashboard that reads the trading bots' live state directly (positions, P&L, recent fills) — no separate demo feed, the stream shows the actual bot state.

ffmpeg captures that virtual display plus audio (routed through a dedicated null PulseAudio sink so nothing else on the machine bleeds into the stream) and tee-muxes a single encode out to multiple streaming platforms simultaneously.

Long-running unattended browser + encoder processes degrade in ways a supervised process never would: Chromium memory creep, a streaming platform's ingest silently freezing, an audio sink quietly unbinding. Each failure mode got its own dedicated watchdog — a memory guard, an ingest-freeze detector, a sink-health check — plus a preemptive daily restart that clears long-uptime Chromium degradation before it becomes visible.

Added a trade-triggered auto-clip pipeline: a rolling ffmpeg buffer is always recording, and a watcher process slices out a clip the moment a real trade event fires — no re-encoding the whole stream, no manual editing — then hands it to an auto-uploader for short-form distribution.

Results

Tech stack

What this means for you

The transferable piece isn't crypto-specific: any 24/7 unattended broadcast, kiosk display, or monitoring wall that has to survive indefinitely without a human watching it needs this same failure-mode-by-failure-mode watchdog pattern — not just "restart on crash," but restart on each specific way this class of pipeline quietly breaks.

The event-triggered clipping pattern generalizes too: anywhere you have a continuous feed but only care about specific moments in it (build-in-public streams, security footage, monitoring dashboards), triggering off the actual event beats a human scrubbing hours of footage or a naive scene-change detector.

Want something like this for your business?

Start with a free 30-min call — figure out fit before money changes hands.