Edson – Python, LLM, RAG

Architected and deployed an agentic aviation maintenance platform for United Airlines, combining a multi-agent ATA-code memory generation workflow, Bedrock-hosted prediction agent, and RAG-based historical case retrieval, reducing technicians' root-cause research time from 2+ hours to ~20 minutes

• Cutting technicians' research time from 2 hours to 20 minutes
• Development of the code base
• Design of the multi-agent architecture

Enhanced a version-aware Hybrid RAG application for United Airlines that identifies context-specific changes across thousands of pages of aircraft manual revisions, reducing pilots' and flight crews' document research time by 60% through semantic + BM25 retrieval, multimodal document processing and RAGAS evaluation (Claude Sonnet on Amazon Bedrock, Milvus, PostgreSQL)

• Reducing document comparison time by 60%
• App validation
• Helping with code development

Built a prompt engineering feature suggesting the next integration component to users (GPT 3.5 Turbo), reducing integration build time by 10%.

• Reducing integration build time by 10%.
• Prompt engineering
• Development of the code base
• Technical discussions involving the architecture

Built ML classification models to predict components at risk of failure, improving maintenance planning by helping crews prioritize high-risk locations. Developed and trained the models using Amazon SageMaker Notebooks and Studio.

• Developed and evaluated ML classification models;
• Collaborated with business stakeholders to clarify requirements and align modeling approaches with business needs;
• Developed unit tests and maintained the codebase in line with software engineering standards.

Built an AIOps system to automatically classify and route support tickets to the appropriate teams, replacing a slow manual process. The solution evolved from classical NLP and ML approaches using BERT, FastText, and Random Forest on Databricks to prompt engineering with Llama 2.

• Developed an LLM-based solution for automated support ticket classification and routing;
• Owned development of the end-to-end codebase supporting the classification solution;
• Deployed and served the model in production using model serving infrastructure.

Ran feature selection to identify the features most driving lead to customer conversion (SageMaker Notebooks).

Deployment of internal web applications

• Discovery of the main features that impact the business
• Development of the code base

Development ETL pipelines using Apache Airflow to populate databases for analytics

• Code development
• Improvement of all ETLs through Apache Airflow.