Manuel – Python, LLM, GCP

The project focused on developing an internal virtual assistant to streamline access to organizational knowledge. The assistant allows business units to upload and query documents using LLMs, enhancing information retrieval and decision-making.

• Implemented a document-based retrieval system using pgvector and integrated multiple LLMs (GPT-4, LLaMA 3.1, DeepSeek) via Azure AI Foundry based on user preference;
• Enabled users to upload and manage various document formats (PDF, DOCX, HTML) within workspaces;
• Communicated feature possibilities, customization options, and tool capabilities to internal stakeholders;
• Monitored and reported performance metrics to inform ongoing improvements.

An AI-powered, voice-guided meditation tool that lets users customize session length and content type. It combines text-to-speech (TTS) technology with soothing background sounds to create a seamless, immersive experience. The platform is designed to support personalized and effective mindfulness practices.

• Conceived and designed the core concept and functionality of the application, aligning it with user needs and mindfulness goals.
• Developed the backend infrastructure;
• Built the front-end interface to ensure a seamless, user-friendly experience.

A global energy company focused on integrating energy transition, balancing traditional energy production with low-carbon initiatives.

• Designed, implemented an end-to-end pipeline and server to receive pdfs and extract structured information from these using object detection and graph neural networks;
• Presented the solution to the business units of the company;
• Assisted them in their initial usage.

The innovation arm of the company, focused on cutting-edge research in areas such as artificial intelligence, wireless communication, and computing. The project involved benchmarking a wide range of Named Entity Recognition models to extract fine-grained categories of entities from news articles.

• Came up with a model XLM-roberta checkpoints combination that gave top performance on a wide range of dataset annotations provided by the headquarters team;
• Worked on the news recommendation features with fine-grained Named Entity Recognition (NER) models;
• Benchmarked Large Language Models (LLMs) for the Huawei HQ news recommender systems team;
• Developed an entity linking solution with sentence similarity models for vector database data ingestion.