user@t8000:/$ ls -l projects/
Apps I built recently, followed by projects delivered to clients.
Sentiment analysis running entirely in the browser. A BERT model is converted to .tflite and the JavaScript loads it from a remote repo, so no data travels to a server.
Any custom model trained in TensorFlow or PyTorch can be converted to .tflite and deployed the same way.
Image classification in the browser with tflite and JavaScript. Instead of the default EfficientNet, I used a custom PyTorch ResNet50, quantized and loaded from a remote repo.
Write-up: In-Browser Inference with a Customized TensorFlow Lite Model and MediaPipe.
Browser-based pose estimation with TensorFlow.js (MoveNet, BlazePose and PoseNet). Users turn on the camera and see real-time skeletal tracking over the video feed, switching between models.
Everything runs client-side: lower latency, no server load, and the video never leaves the device.
An Android app built with an AI code assistant, presented at Pace University, New York, in December 2024.
Three projects at Zendata. The first tests open and closed source LLMs for Security, Brand, Legal and General risk with more than 1,800 prompts. The second extends a code evaluation tool with RAG over a Data Policy document to validate its findings. The third attacks multi-agent systems with prompt attack techniques to measure their robustness.
All three run in production on AWS.
Built in a Google Developer Experts sprint with MakerSuite. The Fitbit Versa 3 manual was parsed with Document AI, served from a Cloud Run instance, and exposed through a Dialogflow CX chatbot.
A Colombian municipality wanted to overcome market information asymmetry with a multiplatform service connecting job seekers and businesses.
We delivered a complete recruitment solution on Google Cloud. The retrieval model matches the top N vacancies to a candidate and the top N candidates to a vacancy. NLP for data preparation, a TensorFlow neural network, daily training with Kubeflow, inference under 100 ms, deployed on Vertex AI. It is an implementation of the Google Cloud Virtual Career Center.
Running at bogotatrabaja.gov.co.
The client needed something better than a decision-tree chatbot that demanded a lot of effort from users.
We hand-built a dataset of laws and chapters and trained a deep network to embed both queries and the database. To keep costs low it runs as a Flask app in a Cloud Run container, called from a Dialogflow CX chatbot that returns the top answer. The client validated it at over 90% assertiveness. More info.
A proof of concept to prevent fraud at the cashier. Existing solutions were not financially viable in Brazil.
We fine-tuned a TensorFlow object detection model on a hand-annotated dataset, first to identify milk cans. A cheap and efficient way to prevent fraud. More info.
A chain with 130 hotels needed to forecast demand beyond its existing 82% accuracy.
Instead of an LSTM we used a Temporal Convolutional Network in TensorFlow, with one-dimensional convolutions over clusters of hotels. The model was pre-trained on stock market data, fine-tuned with transfer learning and retrained weekly by cron on a Compute Engine instance. Average accuracy for the next 3 months: 91.5%.
A government agency was comparing Azure, AWS and Google Cloud for detecting fraud in tax returns.
We loaded the labeled dataset into BigQuery and trained with custom jobs and AutoML, tuning hyperparameters with Vertex AI Vizier. According to the client, our predictions outperformed the competitors.
A trade marketing company needed to cut the cost of hiring large numbers of less skilled employees.
We deployed a virtual career solution in 6 Latin American countries using computer vision, speech analysis, OCR, NLP and semantic similarity of multilingual BERT embeddings. Up to 73% less screening time, up to 78% less time to fill vacancies, over 85% accuracy in choosing candidates. More than 365,000 users by December 2022. The solution was sold to Marco Marketing, Argentina.
A well-known Brazilian coffee master wanted to automate his roasting technique so producers could adopt it.
We instrumented a professional roaster with temperature, pressure and oxygen sensors and adapted the derivative of the roasting curve to improve sweetness, aroma, acidity and body. Deployed on AWS with a human-machine interface, and delivered to one of the largest gas providers in Brazil.
A proof of concept for governments in the Middle East: a recommender that returns the top 10 vacancies for a resume, with the embedding distance of each result. Deployed on Hugging Face Spaces with a Gradio interface.
A 2018 pipeline that streamed CPU temperature to AWS IoT, stored it in DynamoDB and plotted it in QuickSight in real time through Kinesis and Firehose. Details here and here.
In 2018 this was new in Brazil: adding intelligence to call centers. Speech-to-text and NLP extracted data for analytical dashboards, crossed with the customer database. A later version used Random Forests to classify anomalies in call scripts.
A personal swing trading system for the BOVESPA exchange. The Python algorithm combines MACD, moving averages, Fibonacci and RSI to signal buy, hold or sell, and estimates the mood of the market. It runs on Compute Engine on a cron schedule and writes to BigQuery tables that feed a Looker Studio dashboard.