Selected work / Luis Guimarães

Things I built, and how they work.

Each one started as a job someone was doing by hand: copying data between systems, reading through files, checking content before it went out. I designed, built and still run all of them.

Based in Berlin · Registered freelancer in Germany · English & Portuguese

Explore the projects

01 / Portfolio

Five projects. Concrete examples.

My role on each one, how it works, and what you can look at today.

Hermes workflow: selected transcript, two drafts, second review, Telegram selection, and voice and video production — Workflow diagram. Story selection, script choice, and publishing stay with the creator.

Personal workflow

01 / Hermes

AI Content Workflow with Human Approval

I designed and built a workflow that turns a selected video transcript into two draft scripts, runs a second review, and requests a choice through Telegram. The chosen script can move into voice and video assembly. I defined the editorial process and implemented the orchestration, notifications, and approval steps.

My role: Product design and workflow implementation
Skills: AI Automation · Python

Personal content workflow. The creator retains story selection, script choice, and publishing approval.

Workflow & implementation for Hermes

The problem

Content production repeats the same handoffs: choose a source, draft scripts, review them, choose a version, and prepare production.

The workflow

Selected transcript → two script drafts → second review → Telegram notification → human selection → voice and video assembly.

My contribution

Defined the editorial process and decision points; built Python orchestration, document generation, Telegram notifications, and handling of script selections. The deliverable is a repeatable content workflow with documented setup and clear decision points.

Relevant to your project

Draft generation, review queues, team notifications, and workflows that require a human decision before the next action.

Property search in Portuguese with parsed filters, neighborhood price context, source listings and a map of matching neighborhoods — The logged-in search workflow. The live application requires an account.

Live product · Account required

02 / Quem Tem Imóvel

Property Search and Data Integration for Agencies

I designed and built a property search product for Brazilian real estate agencies. It brings sale and rental listings from multiple agency websites into one searchable index. Agents can describe a buyer's request in Portuguese, refine filters, compare neighborhood asking prices, and identify agencies advertising the same property.

My role: Product design, data integration and full-stack development
Skills: Data Integration · Python · Web Development

Live product with account-based access. It helps agents find options beyond their own inventory; no measured sales uplift is claimed.

Open product login

Workflow & implementation for Quem Tem Imóvel

The problem

An agency may receive a buyer's request that its own inventory cannot satisfy. Checking other agencies one by one makes it harder to find suitable alternatives.

The workflow

An agent describes the property request in Portuguese. The application converts supported criteria into filters, searches indexed sale or rental listings, and shows matching properties with links to their source agencies. Agents can refine price and room counts, browse neighborhoods on a map, and see when multiple agencies advertise the same property.

What I built

Product design and a complete web application, including a Python and SQLite ingestion pipeline, source-specific readers, duplicate grouping, query parsing, filters, account-based access, and the search interface. Neighborhood comparisons provide context from indexed asking prices.

Deployment and operations

I configured Docker hosting on my Linux VPS, separate staging and production environments, automated checks before deployment, and releases from a clean checkout matching the main branch. The deployment script tags images by commit, checks application health, and attempts rollback if a release fails its health check. Persistent database volumes keep application data separate from code updates. I also configured scheduled ingestion, offsite backups, and restore verification.

Ordento in Simulator mode: 3D production line with order progress, inspection and rework — Public demo in Simulator mode. All production figures shown are simulated.

Prototype · Simulated public demo

03 / Ordento

Wireless Shop Floor Tracking for Small Manufacturers

I designed and built a product prototype for small manufacturers to follow orders across a 3D shop floor. A QR scanner connected to an M5StickS3 handheld sends operator confirmations over Wi-Fi to an MQTT broker. The backend validates events, stores traceability records in SQLite, and updates the browser. The workflow covers station completion, inspection, rework and scrap, with timing and cost calculations.

My role: Product design, full-stack development and IoT integration
Skills: MQTT · Internet of Things

Self-initiated product prototype. The public demo runs in Simulator mode; physical scanner updates use a separate connected mode. No customer deployment or measured financial savings are claimed.

Explore the simulated floor

Workflow & implementation for Ordento

The problem

Small manufacturers need to know where an order is, whether it is at risk of being late, and what rework or scrap adds to its cost.

The connected workflow

An operator scans a QR code on a route card or command label. An M5Stack U173 reader connects to an M5StickS3 handheld, which sends the event over Wi-Fi using MQTT over TLS. The backend validates the operation, records the event in SQLite, and updates the 3D floor through WebSocket messages. The handheld receives the confirmed state from the broker.

What I built

Product requirements, a React and Three.js floor view, Node.js backend, MQTT integration, MicroPython handheld firmware, printable route cards, station confirmations, inspection outcomes, and rework and scrap records. Timing and cost calculations use the recorded events and configured plans.

CI and VPS deployment

I configured GitHub Actions to run type checks, a production build, and targeted application checks on pull requests and pushes to main. A manually triggered deployment script builds a Docker image from the main branch, tags it by commit, and deploys it to my Linux VPS. It verifies the running commit through the health endpoint and attempts rollback if the new release fails. Persistent storage keeps the application database across releases.

IsleScout property directory: searchable listings across five Greek islands, with location and property filters — A searchable directory bringing property listings from multiple sources into one interface.

Live personal product

04 / IsleScout

Data Integration and Searchable Property Directory

I built a searchable property directory covering five Greek islands. The work combines property data ingestion, normalization and deduplication with location-aware search and filters. A Python and SQLite backend shares query logic between the website and its API.

My role: Product design, data integration and web application development
Skills: Python · Data Integration

Self-initiated product, not a paid client case study. Commercial results are unmeasured.

Explore IsleScout

Workflow & implementation for IsleScout

My contribution

I designed the product and implemented the data pipeline and web application. It demonstrates how scattered source data can become a useful searchable tool.

Relevant to your project

Source integration, record normalization, duplicate handling, and searchable directories with a consistent website and API query path.

Sandkasten search interface: explore Android apps and inspect connected evidence in the knowledge graph — The research interface connects app findings to inspectable source paths.

Personal research project

05 / Sandkasten

Searchable Research Tool with a Neo4j Knowledge Graph

I built a searchable application that connects Android apps, embedded tracking SDKs, their owners, and historical enforcement records. The work combines data ingestion, structured AI extraction, entity resolution, a Neo4j graph, and a web interface with inspectable source paths.

My role: Product design, data pipeline and application development
Skills: Neo4j · Data Integration

Static scan findings show potential connections, not observed data transfers or proof of an app violation.

Explore Sandkasten

Workflow & implementation for Sandkasten

My contribution

Product design, the data pipeline, and application development. The project demonstrates how fragmented sources can become a usable research tool.

Relevant to your project

Search across connected information, resolve entities from multiple sources, and let users inspect the evidence behind a result.

02 / Delivery & operations

From a working prototype to a hosted application.

I deploy and maintain applications on my own Linux VPS. My experience includes Docker, GitHub Actions CI checks, scripted deployments, staging, health checks, backups, and rollback procedures.

Check the change

Ordento runs type checks, production builds, and targeted application checks on pull requests and pushes to main.

Release with control

Manually triggered scripts deploy images tagged by commit. Quem Tem Imóvel has separate staging and production environments.

Keep a recovery path

Health checks, rollback procedures, persistent data, and backup verification support ongoing operation.

03 / More projects

The wider toolkit.

Additional prototypes, product discovery, and infrastructure work.

Working prototype · Small manufacturing

MESkit

A lightweight manufacturing execution prototype with Build and Configure modes, persisted parts and routes, and natural-language shop-floor tools.

Run, monitoring, planning, and MQTT ingestion remain roadmap work.Explore project ↗

Customer discovery · Internal tool

Small manufacturer lead finder

A CLI that searches manufacturing businesses, deduplicates results, checks fit signals, and exports ranked leads for human outreach. It was built to test where MESkit's prospective customers might be found.

No customer wins claimed.

Product concept · PRD

Knowledge Base Graph Chat

A source-cited project archive designed to surface decisions, risks, open questions, and conflicting evidence from files. Its first version specifies a searchable list and grounded chat.

Infrastructure · Local setup

OpenClaw installation & configuration

A documented Docker-based setup for a local assistant environment, including gateway authentication and connection diagnostics.

Configuration work; provider setup and production use are unverified.

Native utility

Free Up Space

A macOS app for finding large or growing folders, with a confirmed, reversible move-to-Trash flow.

Local app; no public time-saved benchmark.

04 / Collaboration

Start with one workflow.

01 / Understand

Map your current steps, tools, volume, and exceptions. Agree what a useful result would look like.

02 / Build a small paid pilot

Define the scope, deliverables, acceptance criteria, and a measure of success. Keep human approval where judgment matters.

03 / Test and hand over

Test agreed examples, document how to run it, and define handoff and any ongoing support before expanding.

05 / Contact

What keeps getting done by hand?

Send me the task your team repeats, the tools you use, and one example of the input and desired result. I can help identify a practical first step.

Get in touch on LinkedIn ↗

← Back to Product Led Lab