Federico – Python, LLM, AWS

This project focuses on building an intelligent agent that interprets user intent from natural language, generates and validates Teradata SQL queries, handles errors with retries, and transforms results into narrative or tabular reports. It simplifies access to complex data, enabling users to get insights without writing SQL.

• Built the Agent source code using Python;
• Set up and maintained the Azure DevOps repository for version control and collaboration;
• Designed and implemented all LLM prompts supporting the Text2SQL functionality;
• Wrote the full report generation logic in Python to automate output based on processed queries;
• Introduced dynamic schema introspection, which reduced runtime errors by 95%;
• Developed and maintained automated test coverage to ensure stability and performance.

A self-hosted Python microservice replicating the OpenAI Assistants/Threads/Runs workflow, including function calls, file handling, and tool routing—designed to ensure vendor independence and full control over the AI interaction pipeline.

• Developed a full-featured replica of the OpenAI Assistants API entirely in Python;
• Created and managed the project repository using Azure DevOps;
• Implemented over 20 API endpoints in line with OpenAI specifications;
• Deployed a private instance of the Llama3 model within a GCP environment;
• Successfully passed a security audit for an on-premise deployment in a regulated industry;
• Built and maintained automated tests to support system reliability;
• Authored a comprehensive 20-page documentation guide covering usage, architecture, and setup.

A baseline retrieval-augmented generation (RAG) service using Microsoft AI Search with hybrid search, followed by a chain-of-thought (CoT) and few-shot prompting to enhance response accuracy and relevance.

• Delivered the foundational setup to enable further research on advanced hybrid approaches;
• Created and configured indexes within Azure AI Search;
• Indexed documents efficiently to support retrieval-based workflows;
• Developed initial Naive RAG implementation using Python;
• Applied optimization techniques to enhance the precision of search retrieval;
• Designed and implemented an evaluation system to track accuracy metrics;
• Documented key engineering decisions, trade-offs, and identified performance limitations.

A robust testing pipeline that auto-generates regression suites for every model snapshot, executes hundreds of prompt–response tests in parallel to ensure consistency across updates, and tracks performance changes by reporting detailed deltas and metrics to Grafana dashboards.

• Developed a PyTest-based testing harness and seed prompt generator, increasing automated coverage from 0 to 85% across known user flows;
• Reduced manual QA efforts from two full days to 45 minutes per release cycle;
• Introduced a statistical analysis layer to distinguish meaningful regressions from noise;
• Connected the system with Azure DevOps, enabling automatic pass/fail badges on all pull requests.

A lightweight add-in seamlessly embedded into Office files, designed to deliver rapid insights by extracting the document structure and generating an executive summary with key action points—all within five seconds. The solution enhances productivity by integrating AI-driven summarization directly into everyday workflows, with minimal overhead and a focus on speed and accuracy.

• Designed the end-to-end plug-in architecture within MS Azure Copilot Studio, including the data privacy flow;
• Indexed documents in Copilot Studio to enhance contextual understanding;
• Configured custom Copilot settings to improve response accuracy;
• Created custom Topics to streamline data flow handling;
• Optimized the summarization pipeline, achieving an average processing time of 3.8 seconds for 10k-word documents.