BTC
$63,973.00
2.51%
ETH
$1,857.35
2.08%
SOL
$73.87
2.92%
XRP
$1.09
2.64%
BNB
$564.13
1.04%
Home/How It's Built

Engineering note

A newsroom with
no journalists

AlfaBrief looks like a crypto-news site. It's really an autonomous pipeline— software that reads the market firehose, decides what actually moves prices, writes the brief, fact-checks its own draft, and publishes. Day to day, no one touches it. This is what's underneath.

2/day
Briefs published
5min
Scan interval
$10/mo
Cost to operate
~60%
LLM cost cut

From firehose to brief, in eight stages

Every stage has one job. The key decision is which stage runs which model — the expensive one only ever sees work the cheap one already approved.

Sonnet — quality model Haiku — cheap model Plain code — no model
01

Monitor

CODE

A handful of crypto news feeds plus a live exchange data stream, checked every five minutes. Everything lands in one queue of candidates.

02

Filter & score — the gate

HAIKU

A cheap, fast model scores each candidate 0–1 for market impact against a fixed rubric. Most are dropped right here. This is the cost gate — it's why the expensive model never has to read the whole firehose.

03

Deduplicate

CODE

A seven-day window drops anything already covered, matched on story identity rather than wording — so the same event never gets processed or published twice, however many outlets run it.

04

Rank & cap

CODE

Survivors are ranked by a composite score — recency, USD magnitude, token significance, source consensus. A hard daily cap then takes only the top few, which bounds both output and the worst-case bill.

05

Rewrite

SONNET

The only place the quality model runs. It turns each surviving story into an original, trader-focused brief — thesis, context, data, what invalidates it — written from the source, not copied from it.

06

Verify — grounding pass

HAIKU

A second cheap-model pass re-reads the draft against the source and strips any claim the source doesn't support. If a rewrite invents a number, this is what removes it before it ships.

07

Fact-check gate

CODE

A deterministic check, no AI in it: every dollar figure in the draft must trace back to the source. An untraceable figure — a made-up ETF outflow, an inflated exploit total — holds the article instead of publishing it. It exists because a soft AI-only check let a fabricated number slip through once.

08

Publish

CODE

The finished brief goes live on the site and into the ~5-day digest — categorised, timestamped, tagged by risk type. A Next.js frontend overlays live prices and market signals from exchange APIs. No manual step anywhere.

The cost engineering

Constraint

One API key. My bill.

The pipeline shares a single LLM API key with another system I run, so every call it makes is money out of my own pocket. Left naive, a busy news day would quietly run up the bill. So the whole design is shaped around spending the expensive model as rarely as possible.

LEVER 1

The cheap model gates the expensive one. A small, fast model filters everything first. The quality model only ever sees the handful of stories that survive — never the raw feed. Reading the whole firehose with the expensive model would cost multiples more for the same output.

LEVER 2

A hard daily cap. However much news breaks, only a fixed number of briefs get the expensive rewrite. The cap turns an open-ended cost into a known, flat one.

LEVER 3

Right model for the job. The rewrite gets the strong model because quality shows. Filtering and fact-checking get the cheap one — that work is judgement, not prose, and judgement is where the cheap model is already good enough.

Net effect: roughly a 60% cost cut with no drop in the quality of what publishes — and a bill that can't run away on a heavy day. The whole system runs for about $10/month.

What I built, and what I'd fix

The system writes the articles. I built the system — and I'd rather be straight about both sides of that than oversell it.

I designed / decide

  • ·The architecture — every stage above and how they hand off.
  • ·The prompts and scoring rubric that define what counts as market-moving.
  • ·The model choices and thresholds, and the cost controls around them.
  • ·The grounding check — deciding a fact-check pass was worth the extra cost.

The system does, unattended

  • ·Reads sources and scores every candidate.
  • ·Writes each brief from scratch.
  • ·Checks its own facts and drops unsupported claims.
  • ·Publishes to the site and digest with no human touch.

Known limits — and what's next

  • No human editor. A deterministic gate now holds any article whose dollar figures don't trace back to the source — the made-up number that used to slip through is caught before publishing. Softer errors (a mischaracterised cause, an overstated trend) still lean on the model, so a human spot-check still matters.
  • Coverage is capped by what's reachable. Some of the best sources block automated access; the pipeline only gets what it can fetch. Next: cross-check a story across two sources before trusting it.
  • Single point of failure. It runs on one server. If that box dies, the brief goes quiet. Next: redundancy and health alerting.

Stack

Deliberately boring where boring is fine, so the effort goes where it matters.

Python · pipelineClaude Haiku · filter + fact-checkClaude Sonnet · rewritesystemd · self-restarting serviceDigitalOcean VPS · always-on hostNext.js + React · frontendVercel · deployExchange APIs · live data

Running autonomously since 27 May 2026 · roughly $10/month to operate.

Get the brief

Digest every ~5 days. Market-moving intelligence for active traders. Free — no spam, unsubscribe at any time.