Felipe Batista

Developed and integrated a custom solution with OpenAI's Assistants API (Beta), leveraging a fully asynchronous, lightly documented aspect of OpenAI's Python SDK.

WORK DONE
• Enabled advanced features such as message streaming, file attachments, and code interpreter usage.
• Architected dynamic vector store selection and implemented an optimized server wrapper using FastAPI and Uvicorn to handle requests at scale.
• Built robust error-handling mechanisms and created comprehensive technical roadmaps for future developments, including dynamic chart plotting, advanced spreadsheet support, and innovative strategies to overcome OpenAI's platform limitations.
• Assisted the development team with several improvements related to the front end, back end, and DevOps.

I developed a full-stack application focused on compliance monitoring and extracting information from PDF documents. Utilized OpenAI (GPT-3.5 and GPT-4), LangChain, and Zod to derive structured outputs from PDFs, which served as the primary driver for the compliance application.

A large-scale evaluation pipeline was implemented, which included acquiring data from multiple sources, tracking experiments, and rapidly benchmarking information extraction using LangChain and Instructor.

I developed cost-effective solutions that employed reasoning and function calling to utilize more affordable models for simpler documents while reserving more expensive models for longer and more complex documents.

A full-stack application was developed using Next.js and PostgreSQL and designs were provided by the client.

I developed a full-stack application that allowed users to chat with multiple documents using state-of-the-art RAG (retrieval augmented generation) pipelines that leveraged both LangChain and LlamaIndex. I implemented a state-of-the-art RAG pipeline that leveraged both a typical PostgreSQL database and a vector database (Pinecone). I also leveraged function calling and other complex LLM techniques to power integrated LLM operations such as browsing, API calls, and additional document retrieval.

Architected and implemented a video processing pipeline to deduplicate videos at scale using deep learning (TensorFlow) to generate video signatures. I also developed a video augmentation pipeline to validate and test deduplication models using OpenCV, MoviePy, and FFmpeg. I implemented several evaluation routines—both visual and quantitative—to evaluate model results.

The project involved the creation of a digital signal processing pipeline to perform non-intrusive load monitoring (a process for analyzing changes in the voltage and current going into a house and deducing what appliances are used in the house as well as their individual energy consumption).

Over the course of the project, I:

1. Reviewed several papers containing the state-of-the-art methods for NILM.

2. Optimized available models to perform an initial POC.

3. Explored different model architectures on a specific setting defined by the client, including:
a. Parallel CNNs
b. Parallel CNNs with LSTMs
c. CNNs with LSTM (bidirectional)
d. CNN with GRU

Models were optimized, and ultimately, the best model was chosen based on the results of a cross-validation routine.

Deliverables were both Jupyter Notebooks and Python Scripts.

Implemented a state-of-the-art image classification pipeline using TensorFlow/Keras.

Tested several different model architectures (including multi-input models with both images and bounding boxes).

Settled on a fine-tuned VGG16 network.

The pipeline included data augmentation, cross-validation, visualization of accuracy and loss across epochs (and sample results).

Implemented a digital signal processing pipeline for freezing of gait detection for patients with Parkinson's.

Replicated state-of-the-art medical, scientific papers using Python, scikit-learn, TensorFlow, Keras, and Jupyter to obtain well-defined baselines.

After in-depth research of DSP techniques for anomaly detection, I implemented a few deep learning model architectures that had never before been applied to this specific domain.

Performed cross-validation to assess model performance, focusing on the model's generalization potential. The model was trained using data from 80% of the patients and tested on the data from the remaining patients.

I used several models in this project, including CNNs, deep ConvLSTM, and parallel CNNs with LSTM.

In order to test the model, I implemented a simple Android application that used the trained model to make a live Freezing of Gait inference based on the phone's accelerometer data. The model was converted to CoreML for future iOS app development.

Other relevant work includes handling class imbalance by adjusting the loss function of the deep learning models, hyperparameter optimization using GridSearch, and simulation of the effect of different window sizes on model performance.