/* AI Engineer specializing in RAG architectures, document intelligence, and LLM fine-tuning. Experienced in building enterprise-grade data extraction pipelines and production-ready APIs */
🟢 Open to work
Currently an Extern @PfizerProjects include AI models and pipelines for extracting structured data and insights from documents, prototyped during coursework and a Pfizer externship.
.jpg)
✅ Verified by Extern
Prototype AI-powered document intelligence with Pfizer—using OCR, LLMs, and RAG to automate real enterprise PDF workflows and build a standout portfolio project.
Fine-tuned a DistilBERT sequence classification model to evaluate community discourse quality by categorizing social media posts into Analysis, Hot Take, or Reaction classes. Developed a comparative benchmarking pipeline evaluating the fine-tuned models performance and inference latency against a zero-shot Large Language Model (LLM) baseline. Preprocessed and tokenized text datasets using Hugging Face pipelines, optimizing training hyperparameters in PyTorch to maximize classification accuracy.
Architected a resilient Flask API using a multi-signal ingestion pipeline to classify text submissions via parallel semantic analysis and native Python stylometric engines (MATTR and Sentence Length CV). Integrated Groq LLM (Llama-3.3-70b-versatile) for semantic pacing evaluations and engineered a Disagreement Override algorithm to shield creators from false positives by routing divergent signals to an Uncertain safety layer. Designed a transaction-safe SQLite persistence layer governed by a strict state-machine workflow (classified, under review, upheld/overturned) and configured Flask-Limiter IP tracing gates (10 req/min).
AI Engineer specializing in RAG architectures, document intelligence, and LLM fine-tuning. Experienced in building enterprise-grade data extraction pipelines and production-ready APIs
I am Oscar Cabezas, a senior Computer Science student at the University of New Orleans focused on artificial intelligence. I have some practical experience and I am working to build stronger technical skills, especially in AI for document analysis and data extraction.
Externships
Pfizer Advanced: AI-Powered Document Insights & Data Extraction Externship
Pfizer
Experience
Senior AI Data Specialist
iMerit Technology · Jan. 2018 10Present
Freelance Geospatial Annotation Specialist
Pathr.ai · Apr. 2023 02Dec. 2025
Undergraduate Machine Learning Research Assistant
University of New Orleans · Mar. 2024 04Aug. 2024
Education
LSU New Orleans (formerly University of New Orleans)
Bachelor of Science in Computer Science, Concentration in Artificial Intelligence and Machine Learning · Class of 2026
CodePath AI Engineering
Applications of AI Engineering & Foundations of AI Engineering · Class of 2026
Delgado Community College
Associates in Business Administration, Graduated with Honors · Class of 2022
Skills
✅ Verified by Extern · ⏱️ In progress
Prototype AI-powered document intelligence with Pfizer—using OCR, LLMs, and RAG to automate real enterprise PDF workflows and build a standout portfolio project.
The project prototyped AI workflows for extracting and structuring data from enterprise PDF documents, combining OCR, large language models, and retrieval-augmented generation. The work surveyed LLM architectures and OCR pipelines, designed a processing pipeline, and produced prototype components for document ingestion and question answering.
I produced a written summary explaining how LLMs tokenize input, use Transformer self-attention to model token relationships, are pre-trained on large corpora, and are fine-tuned with instruction tuning and human feedback.
The submission summarized how LLMs tokenized text, used Transformer self-attention to model token relationships, were pre-trained on large corpora, and then fine-tuned with instruction tuning and human feedback to improve responses.
Fine-tuned a DistilBERT sequence classification model to evaluate community discourse quality by categorizing social media posts into Analysis, Hot Take, or Reaction classes. Developed a comparative benchmarking pipeline evaluating the fine-tuned models performance and inference latency against a zero-shot Large Language Model (LLM) baseline. Preprocessed and tokenized text datasets using Hugging Face pipelines, optimizing training hyperparameters in PyTorch to maximize classification accuracy.
Architected a resilient Flask API using a multi-signal ingestion pipeline to classify text submissions via parallel semantic analysis and native Python stylometric engines (MATTR and Sentence Length CV). Integrated Groq LLM (Llama-3.3-70b-versatile) for semantic pacing evaluations and engineered a Disagreement Override algorithm to shield creators from false positives by routing divergent signals to an Uncertain safety layer. Designed a transaction-safe SQLite persistence layer governed by a strict state-machine workflow (classified, under review, upheld/overturned) and configured Flask-Limiter IP tracing gates (10 req/min).
