PTEN
Back to portfolio
Open Application
Proprietary product, AI curation

Pulso

A daily AI news briefing in Portuguese that publishes itself. Before anything goes live, every link is opened and every summary is checked against the article.

Problem, decision, and evidence

Problem

Most AI news breaks outside Brazil, and waiting for a local outlet to translate it takes days. Automated curation handles the volume, but it publishes with no one checking.

Role and decision

I conceived the product, the editorial rules, and the architecture, and I oversee its operation. The key decision was to put trust in code: the model writes, but text only goes live after checks it cannot get around.

Available evidence

Live since August 25, 2026. By October 7 it had published 88 editions, two a day with no missed day, totaling 1,765 items and no repeated link. These figures describe operations and controls. They do not measure audience or business impact.

Product: the front page with the dated edition, the day's headline, and the model index.
Product: an edition with its summary and priorities. Each item shows the outlet, the newsroom's region, and a link to the original source.
Product: the methodology published for readers, covering what is checked before each edition goes live.
Product: the publisher health page, with the front page age, the retry queue, and the latest runs with cost and failures.
Product: the model catalog with index, price, and speed, each figure with its source and measurement date.

Click to enlarge

Scope of work

Role

Product, editorial rules, architecture, and operations

Audience

Professionals and leaders who follow AI and prefer to read in Portuguese

Methods & technology

TypeScript, OpenAI API, Claude via Bedrock, and AWS

Outcome

AI news from four continents becomes a daily edition in Portuguese, with the source and the checks in plain view

Context

A proprietary product published at dorly.com.br/pulso. It started as a local AI news dashboard and became a public site with dated editions, a searchable archive, topic pages, a model catalog, and an open methodology.

Business problem

Following AI means reading labs and press from several countries every day. Automation handles the volume but creates a new risk. Without human review, a summary can change a number, report as launched what a company only plans to launch, or attribute a quote to the wrong company.

Audience & personas

  • Professionals and leaders who follow AI to make decisions at work
  • Readers who prefer Portuguese and do not want to wait for a local outlet's translation
  • Teams evaluating clipping, market intelligence, or automated AI reports

Adoption barriers

  • Distrust of AI-written summaries without human review
  • Relevant news breaking in another language and another time zone
  • Labs without feeds and outlets that block automated reading
  • Silent failure: the automation reports success while the front page goes stale

Solution design

  • Two inputs: 27 feeds from labs and the press, plus web search for 7 labs without a usable feed
  • Selection by editorial rule: lab announcements sit outside the quota, press coverage is split across Brazil, the US, Europe, and Asia, and each outlet has an item cap
  • Every link is opened before publishing. Anything that returns a 404, lands on the outlet's home page, or carries a different title than promised is dropped
  • A deterministic gate checks money amounts, percentages, precise numbers, and model versions against the downloaded article text
  • A style pass with the Humanizador PT-BR rules, then an audit by a model from another family: OpenAI writes, Claude via Bedrock audits, with one verdict per item
  • When in doubt, the generated text comes out and the item keeps its title, source link, and front-page line. This applies to a failed summary, an unreadable verdict, and an audit skipped by the spending cap
  • An alarm when the public front page stops advancing, and a health page with the front page age, the retry queue, and each run's cost and failures

Scope and deliverables

  • Serverless editorial pipeline on AWS, with scheduling, automatic publishing, and an account separate from production
  • Public site with a front page, dated editions, a page per news item, a paginated archive, search, 12 topic pages, and RSS per topic
  • Model catalog with index, price, and speed, each figure with its source and date
  • Methodology and publisher health pages
  • Auditor evaluation bench with 21 test cases, 13 of them with a seeded defect
  • CI gates for bundle size, accessibility, and checks on the built site

Evidence and signals

  • 88 public editions between August 25 and October 7, 2026, two a day, with no missed day
  • 1,765 items from 24 press outlets and 8 official lab channels, with no repeated link
  • Press coverage split across the US (31%), Europe (27%), Brazil (26%), and Asia (16%). When selection was by recency, the front page came out entirely Brazilian
  • Across the 88 editions, the gate removed generated text from 52 items and the auditor from 12, without taking any item off the site
  • Model cost of about US$0.48 per edition, according to each run's log. It does not include AWS infrastructure
  • Before going live, the auditor caught all 13 seeded defects and rejected no faithful summary. The corpus is synthetic and measures task capability, not family bias
  • No public data on audience or business impact

Lessons learned

In automation without an editor, trust comes from checks the model cannot get around and from limits disclosed to readers. Pulso's strongest bias was in recency sorting, which happened before any model saw the items. No prompt would have fixed that.

Next improvements

  • Take Pulso da Semana from pilot to the automated flow: an audio episode with two synthetic voices, an audited script, and a transcript. The first one, 9 minutes long, is already published.
  • Open a correction channel for readers, with a named owner.
  • Measure the auditor's family bias with summaries generated by the same model that writes the editions.
View implementation details
  • Generation runs in a staging AWS account: Lambda scheduled by EventBridge Scheduler, SQLite for candidates, and Secrets Manager for keys. The public site is static, on the same S3 and CloudFront as this portfolio, and has no public write endpoint.
  • The auditor is Claude Opus 4.6 on Amazon Bedrock, authenticated through IAM, with no third-party key stored.
  • Text from outside, whether a feed, a page, or a search result, is treated as untrusted data. An instruction hidden in an article does not change the outcome: an item whose URL is not in the edition is discarded in code.
  • Seven feeds were removed after measurement. Among them were an outlet that alone took up to 8 slots per edition and a paywalled one whose text could not be checked.
  • Dated editions and news pages stay out of Google's index so they do not weigh on the domain. The front page, archive, topics, models, and methodology are indexable.
  • It took 11 days from the first commit to the first public front page. That figure shows build speed, not operational maturity.

Related services

AI Prototypes & DemosMVPs, automations, and demos to test an idea's value before scaling — using tools like ChatGPT, OpenAI API, n8n, TypeScript, and AWS to prove what works with a tight scope.

Next cases

AI Daily BriefingScattered news becomes a browsable audio briefing with a transcript and sourcesAI Literacy for the Modern WorkplaceFoundations studied at each person's pace and a live discussion built on common ground
AI automation in operation

Do you have an AI workflow that must run without human review of every item?

I can work with your team to design the controls that make that workflow reliable: what the code checks, what a second model audits, what stays with a person, and how much each run may spend.

Discuss an AI workflow