Dragos Dima

I built a retrieval-augmented generation (RAG) chatbot over a large corpus of domain-specific documents, combining state-of-the-art retrieval and reasoning techniques. The pipeline included fine-tuning embedding models on proprietary text data to improve semantic accuracy, and query decomposition to handle multi-faceted or ambiguous user questions efficiently.

KEY TECHNOLOGIES
• PyTorch, LangChain, and OpenAI Embeddings API

HIGHLIGHTS:
• Fine-tuned sentence-transformer models for domain-optimized vector embeddings.
• Implemented hybrid retrieval with re-ranking for context precision.
• Designed a query decomposition module that splits complex user intents into atomic subqueries.

I developed an agentic AI chatbot with dynamic tool use. The system is integrated with a custom MCP (Model Context Protocol) server, enabling secure and structured access to company databases. This enabled the agent to retrieve and update business-critical data in real-time while maintaining strict permission boundaries.

KEY TECHNOLOGIES
• Python, LangChain, Bedrock, and PostgreSQL

HIGHLIGHTS
• Designed and deployed a custom MCP server for controlled database interaction via semantic queries.
• Integrated role-based authentication and logging for enterprise-grade security.

I developed a Named Entity Recognition (NER) pipeline for extracting clinical entities from DICOM metadata and unstructured medical notes, leveraging both fine-tuned language models and synthetic data generation with LLMs. The project focused on overcoming data scarcity and domain-specific annotation limitations by creating high-quality synthetic datasets using prompt-engineered large language models guided by real examples.

KEY TECHNOLOGIES
• Python, PyTorch, GLiNER, LUKE, OpenAI, Hugging Face Transformers

HIGHLIGHTS
• Generated domain-specific synthetic medical text with controlled vocabulary and style for training and augmentation.
• Fine-tuned GLiNER and LUKE models on mixed real and synthetic corpora to enhance recognition of clinical entities such as findings, anatomical locations, and measurement values.

Developed a tool for entity recognition on scanned documents using state-of-the-art Natural Language Processing techniques such as contextualized word embeddings, recurrent neural networks, and conditional random fields in TensorFlow.

Developed a multi-modal deep-learning approach for webpage classification using PyTorch. Among the architectures used in the final product, the most notable was an ensemble of Vision Transformer, Roberta, and Layout LM v3.

I created a keywords generation tool based on the transformer architecture. The encoder was loaded with a pre-trained transformer encoder. It was trained on multiple public datasets and could generate both present and absent keywords from texts.

A PyTorch-based multi-modal architecture for extracting tables and other information (with formats) from PDFs. The architecture was based on Swin-Transformer for the encoder and BART for the decoder. I was involved in pre-training and fine-tuning for the task in question.