Adrian-Constantin Duta

Change Advisory Agent - MVP (Philips) (Lead Role)

• Project Ownership: Independently designed and developed the entire MVP from scratch, leading all phases from architecture to deployment using Gemma 3 LLMs (local/API) and ChromaDB for advanced semantic similarity search.

• Built an agent-based workflow using LangGraph, automating the evaluation of IT change requests for clarity, risk, completeness, effectiveness, and historical consistency.

• Implemented robust data preprocessing pipelines, including normalization and stratification, to improve input quality and analytical accuracy.

• Combined structured field analysis with inter-field and contextual correlation, and retrieved similar past changes via vector search to provide a holistic assessment.

• Delivered results through automatically generated PDF reports, merging structured logic with semantic insights, enabling smarter IT governance.

AI Workbench Platform - Key contributions

• Conducted deep-dive research into “deep-thinking” models that can invoke MCP tools natively within their reasoning loop; identified the open-source Connector-MCP framework and proved its value by building a Wikipedia deep-research agent PoC (FastMCP + ngrok + Anthropic Python SDK). This solution shortens development cycles and removes the need for custom orchestration logic when communicating with MCP servers.

• Created an A2A-protocol integration blueprint that pairs Anthropic Deep Research with Connector-MCP, paving the way for seamless adoption across our AI Workbench platform.

• Built custom Model Context Protocol (MCP) servers on OpenAI, Azure OpenAI, and Anthropic back-ends, deploying them in Windsurf IDE to standardize and accelerate inference workflows.

Agentic Chatbots (Multiple versions of a chatbot designed to identify diagnostics, remediation steps, and issues: either resolving them or escalating by raising tickets)

Developed and iterated on three versions of the chatbot, including two Azure OpenAI-powered web apps, a third improved release, and a LangGraph proof of concept for advanced agentic workflows.

• Played a key role in building two web applications and a third release powered by Azure OpenAI assistants using large language models (LLMs), designed to automate diagnostic and remediation tasks and accelerate issue resolution.

• Led team initiatives during the initial phase of the project, overseeing progress tracking, structured workflow management, and task distribution to ensure smooth execution. Managed key project tasks such as scheduling meetings, tracking milestones, and facilitating early development efforts.

• Provided mentorship and support by explaining AI/LLM concepts and best practices drawn from prior experience on the Codegen project, fostering team collaboration and technical understanding.

• Designed workflows, architectural diagrams, and task-specific assistants, utilizing advanced prompt engineering techniques to optimize assistant behavior.

• Programmatically managed the deployment and configuration of assistants using Python, ensuring seamless integration and improved collaboration between components.

• Built and integrated a testing framework with a simulated user assistant, enabling automated validation of conversations across multiple use cases and user input variations, including edge cases and less structured or ambiguous queries.

• Actively contributed to Streamlit and Azure Web Apps-based implementations, delivering scalable solutions with real-time problem-solving capabilities through chat interactions, which automate tasks and save both time and money

Data Engineering Automation and AI-Driven Solutions

Codegen (Data Engineering Automation Using Generative AI):

• Designed and deployed a Microsoft Teams application that automates the creation of ETL pipelines (e.g., ingestion, curation) and custom solutions based on user prompts, significantly reducing the time and effort required to build data pipelines.

• Integrated Azure Functions, Azure Bot Service, and Adaptive Cards for a seamless frontend experience, while backend processes were built using Python scripts packaged in a wheel and orchestrated with Databricks Jobs/Notebooks.

• Leveraged the Microsoft Teams Developer Kit to enable users to interact with AI agents powered by large language models (LLMs) such as Azure OpenAI, review generated configuration files, and approve pull requests, accelerating project delivery and improving productivity.

• Created and refined statistics tracking systems to collect detailed data from application runs, optimizing performance and supporting the creation of a robust testing framework.