David – Python, LLM, AI

This AI-powered solution deeply analyzed legacy code to extract valuable business rules and logical workflows. It offered insights into legacy system structure, dependencies, and functionality & provided detailed reports and visualizations to accelerate the discovery of legacy systems and the modernization efforts needed.

The main scope of work included the following:

• developed ML models for topic classification;
• designed and implemented ML systems within the CodeMap pipeline;
• orchestrated and managed multiple microservices for seamless integration;
• ran ML tests and experiments to validate performance;
• performed statistical analysis and fine-tuning models based on test results.

This was another internal project for an AI-powered platform that analyzed legacy code: developed GraphRAG that helped teams instantly understand their codebase by mapping file dependencies into intuitive graphs, enabling quick answers to both technical and non-technical questions.

David was tasked with the following:

• designed handlers to manage different code languages;
• created pipelines to ingest codebases over indices;
• created agentic workflows to complete GraphRAG pipeline;
• trained models for jailbreak detection, PII, and secrets masking;
• optimized queries in Neo4j and TigerGraph;
• debugged and resolved application issues;
• wrote unit, integration, and e2e tests;
• integrated with different providers.

This was the internal project for an AI-powered platform that analyzed legacy code: the CodeGen solution enabled organizations to understand and transform legacy code in hours versus months. Leveraging GenAI, CodeGen quickly refactors legacy code into modern programming languages, saving time and money to future-proof enterprise applications.

David's scope of duties covered the following:

• served ML models for jailbreak and PII detection;
• designed ML systems focused on code generation;
• researched and implemented ML algorithms and tools for code creation;
• selected appropriate datasets and data representation methods;
• ran ML tests and experiments;
• performed statistical analysis and fine-tuning using test results;
• trained and retrained systems;
• created LLM chunkers to index codebases;
• designed agentic workflows;
• validated results using an in-house framework.

This was the 2nd project carried out for the AI-powered video messaging platform for the automotive service industry: bot builder AIVA.

David's main contributions included, but were not limited to:

• trained STT models, including Whisper and Nova;
• implemented noise reduction using Krisp;
• developed an NLP engine to extract structured data from text;
• defined the NLP engine flow to retrieve parameters;
• trained TTS models, including Coqui TTS;
• validated responses from OpenAI API;
• integrated components across the system.

This was the 1st project for the AI-powered video messaging platform for the automotive service industry: the platform was integrated with NLP solutions for sentiment analysis using transformers on customer chat interactions, along with AI features like noise cancellation and translations, to enhance communication and customer experience.

Main responsibilities included:

• trained the Bert model with tailored data;
• resolved unbalanced problems in training to avoid false negatives;
• deployed LLM models in AWS;
• monitored and evaluated results over time.