← Back to Work
Work · AI

AI work

Eleven projects I designed and built with AI agents, most of them with Claude Code as my engineering partner. In each one I define the problem, the product and the operating rules; the agents write, test and maintain the system with me.

  1. 01

    Laser Eyes, a trading desk run by AI agents

    Seven AI agents set the allocation of a real crypto portfolio every day, and every decision is published.

    Read the case
  2. 02

    An automated crypto newsroom in three languages

    Selects stories, writes, translates and publishes news in Portuguese, Spanish and English with no editorial staff in the daily loop.

    Read the case
  3. 03

    Autonomous BI with an AI analyst

    Leadership receives the business indicators and relevant anomalies every morning, and an analysis of customer complaints every week.

    Read the case
  4. 04

    AI-assisted sales prospecting

    Detects buying intent in public comments and drafts personal messages, always with human approval before sending.

    Read the case
  5. 05

    FX competitiveness monitor

    Compares the price of a EUR↔BRL corridor with competitors several times a day, by ticket size and payment method. Built in two weeks.

    Read the case
  6. 06

    SEO and GEO agent with more than 400 pages published

    The first production code shipped by someone outside engineering in the company’s history.

    Read the case
  7. 07

    O Vidente, prediction markets on WhatsApp

    Lets users discover markets, deposit via PIX, place bets, track them and cash out through messages in Portuguese.

    Read the case
  8. 08

    An AI chief of staff

    Consolidates email, calendar, meetings, documents and chat every night and prepares the next day’s briefing.

    Read the case
  9. 09

    Local customer acquisition for an optical store

    A site optimised for search and AI, a local SEO strategy, campaigns created through the API and a bot that books eye exams. Delivered in under two weeks.

    Read the case
  10. 10

    B2B lead base for SMEs

    Research agents turned public sources into a qualified, explained company base with a priority list of 50 contacts.

    Read the case
  11. 11

    Forte para a vida, a training programme built as a product

    75 exercises with reference videos, automatically generated GIFs, self-assessments and dose-based routines, based on 33 expert sources.

    Read the case
01 · Runs on its own in the cloud · Started Oct 2026

Laser Eyes, a trading desk run by AI agents

Seven AI agents set the allocation of a real crypto portfolio every day, and every decision is published.

Visit the project · lasereyes.io ↗

Challenge

Investment rules tend to fail in execution. Markets move overnight and discipline gives way to emotion. The AI trading products on offer are usually opaque or run only in simulation.

Solution

  • Five specialist agents (technical, on-chain, derivatives, macro and sentiment) analyse the same market snapshot. A sixth agent challenges the consensus and a chief economist makes the call.
  • A deterministic Python engine turns the decision into orders under more than 20 risk rules, with limit orders only and no withdrawal route in the code.
  • A bilingual website publishes every decision with each agent’s reasoning and performance against holding bitcoin.
  • Failures trigger a maintenance agent that writes the test, fixes the code and ships the fix.

Impact

  • A real-money portfolio managed around the clock with no human input in daily decisions.
  • A decision process the public can audit end to end.
  • Professional desk controls, with more than 800 automated tests, continuous reconciliation and an emergency stop.

Architecture

  1. 25 market sources
  2. Frozen snapshot
  3. 7-agent committee
  4. Risk engine
  5. Limit orders
  6. Public website
  • Python
  • Claude
  • 7 agents
  • sources
  • tests

Screens

Home: the robot’s chart and the squad table with each agent’s stance.
The round table: each agent explains its view; the Chief Economist makes the call.
Launch video: the agents watch the market all the time.
Launch video: every decision is public, with its reasoning.
On a phone.
02 · Runs on its own in the cloud · Started Aug 2026

An automated crypto newsroom in three languages

Selects stories, writes, translates and publishes news in Portuguese, Spanish and English with no editorial staff in the daily loop.

Challenge

The European market lacked a crypto news source for Portuguese, Spanish and English readers. A traditional newsroom in those three languages would need full-time editors, translators and engineers, and existing automated sites merely rewrote other outlets’ articles.

Solution

  • An AI editor reviews candidate stories three times a day and records the reasoning behind each choice.
  • Facts are extracted before writing, and each article is written from scratch with credit to the source.
  • The English version is the reference for translations.
  • Euro prices, an investment calculator, educational content, a newsletter and multilingual SEO.
  • A quality agent monitors the site daily and opens a repair session when it detects a problem.

Impact

  • Daily publishing in three languages, run by one person.
  • Original, credited content with organic search presence in three markets.
  • Continuous quality control that only costs money when something needs fixing.

Architecture

  1. Dozens of sources
  2. AI editor
  3. Fact extraction
  4. Original article
  5. Translation
  6. Site + quality agent
  • LLM
  • languages
  • lines
  • tests
  • ADRs

Screens

Home page with euro prices, lead story and most read.
Prices in euros with 24-hour change and 7-day sparklines.
Bitcoin investment calculator, on a phone.
The home page on a phone.
03 · Runs on its own in the cloud · Started Feb 2026

Autonomous BI with an AI analyst

Leadership receives the business indicators and relevant anomalies every morning, and an analysis of customer complaints every week.

Challenge

Dashboards existed but depended on someone checking them. A drop in the sign-up funnel or in a product could go unnoticed for days, and support conversations were read one by one with no view of trends.

Solution

  • A daily summary in the team chat covering sales by product, inflows and outflows, the sign-up funnel and active customers.
  • Deterministic triggers flag changes, and the AI judges whether they matter before reporting them.
  • On Mondays, a review of the week with deviations, points of attention and hypotheses to investigate.
  • A layer of validated definitions that lets people and agents query the data warehouse safely.
  • A weekly classification of support conversations by theme and trend.

Impact

  • Anomalies handled on the day they appear.
  • The team starts each week with the main customer pain points prioritised.
  • A process with no manual steps and no dependence on a dedicated analyst.

Architecture

  1. Data warehouse
  2. Product analytics
  3. Support platform
  4. Triggers + AI analyst
  5. Team chat
  • Python
  • SQL
  • LLM
  • Data warehouse
  • serverless scheduler

Screens

Internal system with no public screens.

04 · Runs on its own in the cloud · Started May 2026

AI-assisted sales prospecting

Detects buying intent in public comments and drafts personal messages, always with human approval before sending.

Challenge

Remittance customers look for the service only when they need it, and paid media reaches them at a high cost. Meanwhile, these people were describing problems with transfers and fees in public community pages, and nobody followed those conversations in a structured way.

Solution

  • Comments are collected from public profiles, with each author logged as a lead and the comment kept as context.
  • The AI rates intent, priority and risk and suggests the next step.
  • Personal messages built around the problem described, with no links in the first contact.
  • An approval panel where the operator approves, edits or rejects every message.
  • Operator corrections feed continuous improvement of the messages.

Impact

  • New customers from an audience paid media was not reaching.
  • A manual routine reduced to one approval per message.
  • Continuous operation with automatic reconnection and protection against duplicate sends.

Architecture

  1. Public comments
  2. AI qualifies
  3. AI drafts
  4. Human approval
  5. Paced send
  6. Learning loop
  • TypeScript
  • LLM
  • queue
  • Speech-to-text
  • human in the loop

Screens

From public comment to approved message: collection, commercial AI, human gate, real send.
Built to run in production: auto-reconnect, idempotency, tracking and a learning loop.
The operator cockpit, on the simulation bench.
05 · Runs on its own in the cloud · Started Feb 2026

FX competitiveness monitor

Compares the price of a EUR↔BRL corridor with competitors several times a day, by ticket size and payment method. Built in two weeks.

Challenge

Price checks were manual, one amount and one moment at a time, with no history. Since competitors charge differently by pay-in and pay-out method, there was no way to measure real competitiveness or estimate the effect of a fee change.

Solution

  • Collectors quote the product and competitors at eight ticket sizes per direction and across every method, with the central bank reference rate and a screenshot of each quote.
  • A historical simulator showing the median price and win rate for any amount and period.
  • A fee simulator that re-prices history and estimates the effect of a change before it is made.
  • On-demand quotes and a health panel for every collection run.

Impact

  • Competitiveness measured by ticket size and by method.
  • Pricing decisions based on history and evidence.
  • The effect of fee changes estimated before they go live.

Architecture

  1. Product site
  2. Competitors
  3. Reference rate
  4. Cloud collectors
  5. History + evidence
  6. Dashboards & simulator
  • TypeScript
  • headless browser
  • managed DB
  • serverless scheduler

Screens

Historical competitiveness simulator.
EUR→BRL corridor view: spread against the reference and winning method.
Fee what-if: automatic takeaways and the per-bucket table.
06 · On demand · Started Mar 2026

SEO and GEO agent with more than 400 pages published

The first production code shipped by someone outside engineering in the company’s history.

Challenge

Search demand for buying digital assets far exceeded what the website covered. Pages were produced by hand by a small team, which made one page per asset in three languages unfeasible. AI search was emerging as a new channel with no established method.

Solution

  • Keyword and competitor research, a topic map and a study of visibility in AI answers.
  • One content standard per asset, applied to more than 400 assets in three languages by sub-agents working in parallel.
  • Technical SEO fixes to the sitemap, robots, llms.txt, meta tags and duplicate URLs.
  • A pipeline that runs research, strategy, content and visual review, then opens a pull request for engineering approval.

Impact

  • More than 400 asset pages in production, in three languages.
  • Content structured for traditional search and for AI answer engines.
  • A new model of collaboration between business and engineering, with reviewed code delivered by the business team.

Architecture

  1. Search data
  2. Data-rich brief
  3. Agent with skills
  4. Validation + visual QA
  5. Pull request
  6. Website
  • Python
  • LLM agent
  • SEO data API
  • Pull requests
  • languages

Screens

Catalogue of generated landing pages, by asset category.
07 · Runs on its own in the cloud · Started Jan 2026

O Vidente, prediction markets on WhatsApp

Lets users discover markets, deposit via PIX, place bets, track them and cash out through messages in Portuguese.

Challenge

Prediction markets run in dollars, in English and on crypto infrastructure. For a Brazilian user, taking part requires an exchange, stablecoins, a wallet and familiarity with an order book. WhatsApp, in turn, offers no charts or reliable buttons for a product this heavy on numbers.

Solution

  • A persona with a clear proposition and numbered menus answered with a single digit.
  • Prices shown in reais, such as “YES: R$10 → R$39”, instead of probabilities.
  • Natural-language orders with explicit confirmation before any money moves.
  • PIX in and out, automatic redemption of winning bets and a daily summary.

Impact

  • From deposit to an open position in a few messages, using only WhatsApp and a banking app.
  • Intent accuracy raised from 20% to 98% in one morning of testing and held in place by 148 regression cases.

Architecture

  1. WhatsApp
  2. Intent AI
  3. Conversation engine
  4. Prediction markets
  5. PIX
  • TypeScript
  • LLM
  • WhatsApp
  • PIX
  • lines

Screens

“I am the Seer”: welcome, menu and markets with an R$10 payoff simulation.
Natural-language order, confirmation and the executed bet.
“Your bets”: the portfolio in one message.
QA rig: intent accuracy from 20% to 98% in one morning.
08 · Scheduled on the user’s machine · Started May 2026

An AI chief of staff

Consolidates email, calendar, meetings, documents and chat every night and prepares the next day’s briefing.

Challenge

An executive’s working context is spread across five tools with no integration. Preparing for the next day required manual searching, and nothing linked what was discussed in meetings with what arrived by email.

Solution

  • A single knowledge base, with every source stored as a dated, versioned file.
  • A nightly synthesis with updated pages on people and projects, a pattern report and the next working day’s briefing.
  • Questions from the phone, answered with sources.
  • External invitations received by email added to the calendar automatically.

Impact

  • Daily preparation reduced to a quick-read briefing.
  • Decisions and follow-ups tracked across email, chat and meetings.
  • Data kept exclusively on the user’s machine.

Architecture

  1. Email, calendar, meetings, docs, chat
  2. Versioned knowledge base
  3. Nightly AI synthesis
  4. Briefing + questions from the phone
  • Python
  • Claude Code headless
  • git
  • text search

Screens

Internal system with no public screens.

09 · Cloud with an operator · Started Mar 2026

Local customer acquisition for an optical store

A site optimised for search and AI, a local SEO strategy, campaigns created through the API and a bot that books eye exams. Delivered in under two weeks.

Challenge

The store competed with dozens of opticians in the same area, with no website and little visibility in the most relevant local searches. Competitors were already buying Maps ads, and paid clicks rarely turned into visits.

Solution

  • A diagnosis covering city search volume, Maps rankings and an analysis of 32 competitors.
  • An 11-page site with structured data and answers prepared to be quoted by AI engines.
  • Three Google Ads campaigns created by script, all pointing to the free eye exam.
  • A WhatsApp bot that books the exam, sends reminders and guides the customer to the store.
  • Processes documented as playbooks any AI agent can run.

Impact

  • More store visits and sales, with paid clicks converted into booked exams.
  • A web presence built for traditional search and for AI search.

Architecture

  1. Search & Maps data
  2. AI agent + playbooks
  3. Site + ads
  4. WhatsApp bot
  5. Store visit
  • Astro
  • Google Ads API
  • Claude Code
  • Grok
  • WhatsApp

Screens

Service page with the quick-answer block for AI engines.
GEO baseline: do AI answers cite the store when someone asks for a recommendation?
Step-by-step Google Ads guide generated for the owner.
The site on a phone.
10 · On demand · Started Mar 2026

B2B lead base for SMEs

Research agents turned public sources into a qualified, explained company base with a priority list of 50 contacts.

Challenge

The sales team needed a qualified list of small and mid-sized companies. Bought lists offer volume without fit, and building one by hand takes weeks, rarely with explicit prioritisation criteria.

Solution

  • More than 18 public sources and about 1,800 companies analysed.
  • Collection of website, contact channel, decision-maker role and signals of size and internationalisation.
  • Three scores from 1 to 5, a final priority with a confidence level and a rationale for each selected company.
  • Delivered in a CRM-ready format, with a report on the 50 priority companies in eight groups.

Impact

  • A more focused sales approach, leading to new sales, customers and partnerships.
  • Weeks of manual work completed in one morning, using only public sources.

Architecture

  1. 18+ public sources
  2. Research agents
  3. Scoring & tiers
  4. CRM dataset + report
  • research agents
  • scoring model
  • PDF report

Screens

Top 50 for immediate approach, by tier (companies blurred).
11 · On demand · Started Oct 2026

Forte para a vida, a training programme built as a product

75 exercises with reference videos, automatically generated GIFs, self-assessments and dose-based routines, based on 33 expert sources.

Challenge

Years of machine-based training led to pain and stiffness. Generic programmes treat symptoms without explaining the purpose of each exercise, and quality knowledge is scattered across books, podcasts and videos.

Solution

  • A root-cause diagnosis, warning signs and 16 self-assessments.
  • The same plan in six durations, from 16 minutes a day to a full week, with routines before and after padel.
  • A library of 75 exercises with video, GIF, dose and objective progression criteria.
  • A static web app with no dependencies that opens instantly on a phone.

Impact

  • A complete programme that can be consulted at the gym or on court.
  • Routine changes made in minutes through a simple data edit.

Architecture

  1. 33 sources
  2. AI synthesis
  3. Structured data
  4. Auto GIFs
  5. Web app
  • HTML/CSS/JS
  • Claude Code
  • yt-dlp + ffmpeg
  • Playwright

Screens

Home: “root first, then load”.
Workout A: executive summary with GIFs and doses.
On a phone.