Close

AI Workflows & Automation

LLM content pipelines · AI agents with human sign-off · Scheduled jobs · Bots · Prompt frameworks

I build pipelines where AI does the repeat work and a person signs off on what goes out.

El Mercado

Prediction market · 2026 · Founder

Prediction market of my own, 18 background jobs and an AI news watcher.

24 hours a dayIt runs itself

18 background jobs open the markets, price them, resolve them and pay out.

every hourFixturesEach match opens at the price the odds imply.
every 15 minWeather stationsWill it rain tomorrow, city by city.
every dayThe dollarAgainst the guaraní.
every 5 minBitcoin windowsHigher at 07:00 than at 06:55? Open, close, pay.
18 background jobsOpen, price, resolve, pay outWhen the result is in, the market pays. Nobody touches it.
house accountAlways someone on the other sideIt keeps offers on every market.
since launch26,000+ markets resolvedwith no manual work
A five-minute Bitcoin window closes and resolves.

Every morningAI watches the board

Every morning an internal AI model reads the news and emails me when one of them has already happened.

the gapMarkets with no feedA minister leaving, a bill passing. Someone has to call it.
every morningThe watcher wakesTakes every open market I resolve by hand.
tavilySearches the newsTwo weeks back, Paraguayan outlets first.
geminiHas this already happened?A yes only counts with a quoted sentence. A no leaves a receipt and it tries again tomorrow.
email + adminI get the quote and the sourceThe market moves to the top of the admin.
meI decideThe AI never closes or pays anything.
Email from the watcher: the quote and the source
Watcher email example.

6 monthsResults

26,000+markets resolved with no manual work
623commits
80working sessions with Claude Code
18scheduled jobs
52written specs
97test files

The product works and pays out on its own. A semi-autonomous model.

See the full case →

Matiklava Goes Partying

Weekly newsletter · 2026 · Founder, editor and builder

Newsletter pipeline of my own, each issue drafted for my review.

Matiklava Goes Partying is a weekly guide to underground parties in Taiwan, started in August 2024 for 30 people. Every issue used to take three to four hours of copying flyers into MailerLite. In April 2026 I rebuilt it using Claude Code.

April 2026Pipeline

A script reads 55 Instagram accounts and Resident Advisor, checks the dates and links, writes a draft issue and asks for a manual review.

01 meKickoff“Let's start issue #27.” Eight questions, then it works alone.
02 scrape55 Instagram accountsVenues, collectives, promoters. Resident Advisor cross-check.
03 extractLLM reads every postCaption and flyer become an event: date, venue, lineup, price, links.
04 checkValidatorsWrong dates, dead links, missed posts, Gmail's 102 KB clip limit.
05 buildEmail + archive draftFlyers to Cloudinary, MailerLite draft, archive page, IG carousel.
06 meReview and sendRead the rendered issue, cut what doesn't belong, hit send.
out800+ inboxesand the Telegram bot pings its subscribers
outArchive pagearchive.matiklava.com, a week later
outInstagram carouselfive slides from the same data

Every weekAI-assisted direction

  • "Let's start issue #27."

  • It asks me eight questions about the current issue.

  • It builds the issue and checks it against the original posts.

  • It hands it to me to read.

  • Every correction I make becomes a rule, so the same mistake doesn't come back.

Pythonthe pipeline
Instagram + RA55 accounts, cross-checked
LLMcaption and flyer to event
MailerLitedrafts and sending
Cloudinaryflyers that never expire
Cloudflare Workersboth Telegram bots

24/7Two Telegram bots

Both on Cloudflare Workers: one reports to me, one serves readers.

triggerA pipeline step finishesScrape, cross-check, generate, link check, publish.
cloudflare workerOperator botReports each step. Answers /subs, /trends, /burn.
meMy phoneI stopped watching the terminal.
triggerAn issue goes outOr a reader types /weekend, /next, /venue.
cloudflare worker · launching soon@MGPguideBotParties by day, a card per party, tickets and map.
readersIts subscribersPinged when a new issue is out.
The reader bot.
Telegram chat with the operator bot: scrape, cross-check, generate and link-check reports
The operator bot.
Operator bot commands.
1 / 3
See the full case →