Case study · autonomous publishing system

Keeping up took 14 hours a week. Now it takes 12 minutes.

A machine that watches, reads, and publishes itself.

I follow two crypto analysts whose real analysis lives in multi-hour livestreams. Staying current was a part-time job. So I built a system that does the watching — it finds every new episode, transcribes it, distills the week into one dated synthesis, and publishes the result to this site on its own. This page is that system, shown in public.

01
The problem & the result

The math didn't work by hand

Before
~14 hrs / week

of livestream across just 2 channels — episodes averaging 2.4 hours, the signal buried in the noise.

After
12 min read

one synthesis a week — what they think, where they agree, where they flatly disagree. Every claim dated and traceable to its episode.

~70×
time compression
78
episodes analyzed
2
channels tracked
weekly
self-refresh
02
What I built

A pipeline that does the watching

It discovers each new episode, transcribes it, and analyzes it into a structured report. Those reports compound into a living body of knowledge, and once a week the system distills all of it into the synthesis you can read here — a global market read plus a per-channel dashboard. No human copies anything by hand. The summary is the visible part; keeping it trustworthy is the real work.

READ

This week's synthesis

The global read — consensus and conflict.

RANK

Conviction board

A ranked, gated view of the standing calls.

CREDIT

The channels

The creators whose work this is built on.

03
Under the hood · for engineers

The hard part is keeping an AI system trustworthy

The design principle: a deterministic CLI owns every mechanical step, and an agent is invited only where genuine judgment is needed — always inside a versioned contract, always verified afterward. Determinism where you can, judgment where you must, a check on both.

Deterministic core
Every mechanical action is one reproducible command in a Python CLI — scriptable and testable, no hidden state.
Judgment, on rails
The agent's inputs and required outputs are pinned in versioned contract files — judgment happens inside a fixed shape, not a free-form prompt.
Verified transitions
State only advances when a run checks out Expected vs Claimed vs Actual. Mismatch means refuse — the pipeline fails closed.
The pipeline · 6 stages
1 Scan Discover new episodes per channel (YouTube RSS / Rumble). deterministic
2 Transcribe Fill the transcript (caption API or local speech-to-text). deterministic
3 Intake An agent analyzes one transcript into a structured report. agent · on rails
4 Compound Coins, themes, frameworks accumulate into a knowledge base. deterministic
5 Synthesize The weekly pass distills it into the global + per-channel reads. agent · on rails
6 Publish A deterministic gate renders only the safe pages to this site. deterministic

Full architecture & trust model →