In this Google Developers Experts sprint, the idea was to create a Generative AI application using MakerSuite. 

First, I prepared the Fitbit Versa 3 manual in Google Cloud Document AI and translated into a text that was inserted in a Python notebook. This Python notebook was placed in a Cloud Run instance (serverless) and called the Generative AI application for text generation. The interface used was a chatbot in Dialogflow CX.

The project is located in my GitHub

In this project, a colombian municipality was looking for ways to overcome the market information asymmetry and create a multiplataform solution to connect job seekers and businesses. 

We delivered a complete recruitment solution using Google Cloud technologies, following best practices for the Web application security and also for the Machine Learning pipeline. The retrieval algorithm is responsible for matching top N vacancies to a given candidate, and also suggest top N candidates for a specific vacancy. A Natural Language Processing was applied to prepare raw data, the neural network was developed in Tensorflow and the pipeline for daily training was made with Kubeflow, with an inference time smaller than 100 ms. This solution was deployed in Vertex AI. This project is an implementation of the Google Cloud VCC (Virtual Career Center).

In this project, the client wanted to get a better option than his current decision-tree chatbot solution, that required lots of user effort to work properly. 

We approached the problem with a retrieval recommender. A dataset was made by hand including the laws and their specific chapters and a deep neural network was developed to generate embeddings for the query and for the database. Due to costs, the solution was deployed in a container in Google Cloud via Cloud Run and a Flask application. The chatbot was developed using Dialogflow CX and, given the user question, an endpoint in Cloud Run was called, providing the best answer (TOP 1 candidate). The solution was validated by our client, with more than 90% assertiveness and made easier for citizens search for answers to their requests. More info here.

In this Proof of Concept, supermarkets were looking for ways to prevent fraud at the cashier. Existing solutions were not financially viable for Brazil, due to the low cost of products and wages of cashiers. 

We developed a object detection in Tensorflow using a custom dataset annotated by hand and fine-tuned the deep learning model. The model was initially developed to identify milk cans and generated an efficient and cheap way to prevent fraud. More info here.

A hotel chain with 130 units needed to forecast demand beyond their 82% accuracy, in order to increase their management abilities and generate more profitable marketing campaigns. 

Working with the on-premises database team and with the BI team, we gathered time series data for our model. Instead of a LSTM neural network, we opted for a Temporal Convolutional Neural Network in Tensorflow, with one-dimensional convolutions for clusters of hotels. The model was trained on the stock market, fine-tuned with Transfer Learning, and was deployed in a Compute Engine instance given the budget constraints and performing a weekly training job using cron. Following BI team demands, data was exported as a .csv to serve as input for the Dashboards. The average accuracy achieved for the next 3 months forecast was 91.5%. We didn't use Prophet or ARIMA due to high inference time.

A government agency was comparing different cloud solutions (Azure, AWS, Google Cloud) to detect fraud in citizens' tax return. They needed to automate the process in order to increase their action scope.

We uploaded the labeled dataset to BigQuery, and generated different trainings, with custom jobs and Google Cloud AutoML, optimizing hyperparameters with VertexAI Vizier. We delivered the predictions and, according to the client, our predictions outperformed our competitors.

Our client, a trade marketing company needed help to reduce the costs of hiring a great amount of less skilled employees.

We deployed a virtual career solution in 6 countries of Latin America, that uses computer vision, speech analysis, OCR, Natural Language Processing and semantic similarity of embeddings (BERT multilingual) with cosine distance. As a result, recruiters got up to 73% of average screening time reduction in selection processes, up to 78% of vacancies replacement time reduction and more than 85% of accuracy in choosing candidates. As of December 2022, the platform was used by more than 365,000 people.

A well known brazilian coffee master was looking for a way to automatize his technique to roast coffee to ease the adoption by coffee producers.

We added sensors to a professional coffee roaster to collect temperature, pressure and oxygen. The derivative of the roasting curve was calculated and adapted to generate a coffee with better properties, as sweetness, odor, acidity and body, according to the glucose levels during and after the roasting process. The project was delivered to one of the biggest gas providers in Brazil. The solution was deployed in AWS and we also delivered a HMI (Human Machine Interface).

Our prospects (governments of Middle East) were looking for technology to increase the hiring process efficiency.

We developed a Proof of  Concept of a retrieval model for a career solution, deployed in HuggingFace Spaces using Gradio interface. The model is a recommender system that brings TOP 10 vacancies according to the resume of the candidate, and also the distances of each result according to semantic similarity of embeddings. We used a limited dataset for this PoC.

IoT Streaming Solution in AWS

In 2018 I developed a streaming IoT solution that collected CPU temperature data, streamed data to AWS IoT, saved in DynamoDB and used Kinesis and Firehose to plot temperature data in QuickSight in real time. Project details  here and here.

Back in 2018, this solution was innovative in Brazil.  The idea was to add insights and intelligence to call centers, beyond their current technology.

We used Speech-to-Text and Natural Language Processing algorithms to extract useful data for analytical dashboards. Voice data was then crossed with customer database data to generate additional insights. In a later development, the solution added Random Forest algorithms to classify anomalies in scripts, using voice data.

The interactive dashboard below was made with Looker Studio for personal purposes. The Python algorithm considers financial indicators as MACD, Moving Averages, Fibonacci and  RSI (Relative Strength Index) for BOVESPA stock exchange and forecasts compra (buy), espera (keep) and venda (sell) according to the rotation graph. The mood of the market is also calculated and by applying filter according to the forecast, best suggestions come in green (buy) and red (sell). The Python algorithm was deployed in Compute Engine with a cron job, instance schedule and the outputs are written in BigQuery tables on Google Cloud, that feed the dashboard. Data updated on weekdays, for swing trading.