Lost Threads 2: The Quickening

Last year, I documented a journey building a system to find hidden moments of human vulnerability on Reddit.

You can read about it here: https://jeffcove.com/2025/10/11/lost-threads-a-gentle-introduction.html

I have quietly and consistently been putting in work on this. I want to talk first about how the system has improved me, and then the improvements I made to it.

The core of the work on me occurred through an extensive amount of human-in-the-loop refinement for my analyzers.

Describing an algorithm that finds vulnerability amongst noise forces you to understand its surrounding context. This is a working theory of poems. For example, “I am alone" has no charge on its for exactly the same reason it's inert to the analyzer. The line only hits when everything around it has been quietly building toward it.

I’ve always had an instinct for this but now I know it mechanically, because I had to explain it to a machine that believes nothing by default.

I’ve also learned more about the shape of human discourse. I found that once a thread tilts towards politics or war, it often does not recover to meaningful discussion about anything else. I added circuit breakers to find these instances.

On the technical side:

• I built an entire homelab of cheap mini-PC’s, two of which run LLMs for the final classification steps.
• I re-architected the entire system to use a homebrewed NATS messaging backend so all analyzers use events, send broadcasts, etc. I have the whole thing observed through OpenTelemetry.
  

It runs 24/7 and costs me only $2/week in electricity.

The machine reads all night. In the morning, I read what it found.