Marcelo – Python, Tensorflow, Pandas

A modular, multi-provider LLM platform designed to enable real-time conversational AI with dynamic model switching. It was built to address key limitations in traditional chatbot systems, including vendor lock-in, high operational costs, and limited flexibility. It was designed as a reusable AI platform for client-facing solutions, including customer support assistants and domain-specific chatbots, enabling rapid adaptation across industries.

• Designed and implemented end-to-end system architecture (frontend, backend, and cloud infrastructure);
• Built a multi-provider LLM abstraction layer supporting OpenAI, AWS Bedrock, and Ollama;
• Implemented real-time streaming responses using Server-Sent Events (SSE);
• Developed a scalable FastAPI backend for handling conversational workflows;
• Built a React-based chat interface with dynamic streaming UX;
• Deployed and managed services across Vercel, Render, and AWS (ECS, EC2, ALB);
• Resolved production issues related to CORS, networking, and environment configuration;
• Designed the platform for reusability across multiple client use cases.

An automated ICD-10 coding system for medical reports using Large Language Models and Retrieval-Augmented Generation (RAG), designed to reduce manual workload and improve coding consistency in clinical workflows. The system processes unstructured medical reports and generates structured ICD-10 codes by combining a Claude-based LLM (via AWS Bedrock) with a Knowledge Base containing relevant medical context. This approach ensures more accurate, context-aware predictions than standalone LLM prompting.

• Designed and implemented an end-to-end automated ICD-10 coding pipeline using LLMs;
• Integrated AWS Bedrock (Claude) with a Knowledge Base to enable RAG-based predictions;
• Built a scalable serverless architecture using Lambda, Step Functions, and API Gateway;
• Developed logic to process and structure unstructured medical reports;
• Designed and managed data persistence using Amazon RDS;
• Optimized prompt engineering strategies to improve code accuracy and consistency;
• Ensured system scalability for processing large volumes of clinical data;
• Collaborated with domain requirements to align outputs with medical coding standards.

An LLM-based tool to detect and correct addenda in medical reports, aimed at preventing errors before reports are delivered to patients or external systems. The solution was designed to act as a validation layer within the reporting workflow, helping clinicians reduce the risk of sending incorrect or incomplete information.

• Designed and implemented an LLM-based validation system for medical reports;
• Developed prompt engineering strategies to detect inconsistencies and errors;
• Integrated AWS Bedrock (Claude) for context-aware analysis of clinical text;
• Built logic to identify addenda and suggest corrections in real time;
• Focused on minimizing false positives in a high-sensitivity environment
• Designed the system to integrate into existing clinical workflows;
• Contributed to improving report quality and reducing the risk of human error.

An LLM-based system to automatically classify medical reports as normal or abnormal, to optimize storage, prioritization, and clinical workflows. The solution was designed to reduce storage costs and improve operational efficiency by identifying reports that could be safely archived or deprioritized after a defined period. Special attention was given to prompt design and output consistency to ensure reliable binary classification in a high-sensitivity healthcare context.

• Designed and implemented an LLM-based classification system for medical reports;
• Integrated LLAMA 3.1 using LangChain for structured text analysis;
• Developed prompt engineering strategies for consistent binary classification;
• Built scalable pipelines for processing large volumes of unstructured data;
• Optimized classification reliability in a high-sensitivity healthcare context;
• Contributed to reducing storage costs by identifying non-relevant reports;
• Integrated the solution into existing data processing workflows.

A computer vision system for detecting lesions in mammograms, including nodules, calcifications, and microcalcifications, to support early breast cancer detection. The solution was designed to assist radiologists by highlighting suspicious regions, reducing the risk of missed findings, and improving diagnostic workflows.

• Built and trained YOLO-based object detection models for medical imaging;
• Designed the full pipeline from dataset preparation to model evaluation;
• Worked with multiple YOLO versions (v5, v8, v9) to compare performance;
• Addressed challenges related to detecting small and low-contrast features;
• Validated and refined annotated datasets to improve model quality;
• Optimized model performance through hyperparameter tuning and experimentation;
• Developed region-based detection strategies for complex medical patterns;
• Contributed to improving diagnostic support tools for radiologists.