Zisis – Python, LLM, LangChain

The role involved acaling an AI Engineering program requires more than just technical content; it demands a cohesive academic strategy, consistent mentorship quality across a global team, and the integration of rapidly evolving GenAI tools into a stable educational framework for over 500 running students.

• Designed the comprehensive curriculum and academic strategy for the AI Engineering course;
• Managed and mentored a team of 15 AI instructors and mentors, overseeing both operational and academic quality;
• Built a custom GenAI internal toolset (FastAPI, PydanticAI, LangChain, n8n) to automate student progress tracking and feedback;
• Provided technical consulting and mentorship to over 200 students on their capstone AI projects;
• Oversaw the integration of cutting-edge technologies like PydanticAI and agentic workflows into the teaching material.

Large-scale research initiatives require the coordination of multi-stage AI pipelines that achieve high resource efficiency while meeting strict latency requirements. An AI Orchestrator was architected based on Apache Airflow to automate the end-to-end lifecycle of machine learning models. A key technical focus was advanced model distillation, in which highly efficient Small Language Models (SLMs) were derived from massive LLMs to perform specialized tasks such as cybersecurity anomaly detection. This system streamlined the delivery of AI components across international research consortia, reducing operational overhead and enabling the deployment of AI models on resource-constrained edge devices through optimized distillation.

• Architected a modular AI Orchestrator using Apache Airflow to handle complex, multi-stage research workflows;
• Implemented advanced model distillation techniques to create high-performance Small Language Models (SLMs);
• Managed technical deliverables and cross-functional communication;
• Containerized the entire AI stack using Docker to ensure reproducible environments across different research nodes;
• Designed automated data generation pipelines to facilitate training in data-scarce cybersecurity domains.

The project involved fine-tuning of Large Language Models to achieve high-level emotional intelligence, grounding the training process in social and emotional psychology theories. Using Reinforcement Learning from Human Feedback (RLHF) via DPO and GRPO techniques, the models were optimized for both "emotional understanding" and "emotional application". As a result, developed models capable of identifying subtle toxicity and engaging in empathetic dialogue, which were successfully trialed for content moderation in high-profile social environments.

• Led the fine-tuning of Large Language Models using SFT, QLoRa, and DPO to embed emotional intelligence;
• Configured and managed multi-GPU training environments using FSDP (Fully Sharded Data Parallel);
• Implemented Reinforcement Learning from Human Feedback (RLHF) loops to align model responses with psychological theories;
• Developed evaluation frameworks to measure model empathy, toxicity, and grounding in social psychology;
• Integrated multimodal RAG capabilities to allow models to process and respond to emotional cues in diverse data formats.

A multi-agent system using LangGraph and Autogen to automate the creative workflow. The system consists of specialized agents (Strategist, Copywriter, Critic) that collaborate via an autonomous agentic flow to generate campaign assets. The Model Context Protocol (MCP) was implemented to allow these agents to access real-world data and external tools securely and efficiently. This reduced content production time from days to minutes, enabling real-time ad generation that aligns with brand voice and specific psychological triggers.

• Architected a multi-agent system using LangGraph and Autogen to automate the creative advertising lifecycle;
• Developed specialized agent personas with distinct reasoning paths for strategy, copywriting, and quality assurance;
• Integrated the Model Context Protocol (MCP) to provide agents with real-time access to market data and external tools;
• Optimized agentic flows to reduce token consumption while maintaining high-quality creative output.

A suite of LLM-driven solutions focused on e-commerce personalization and predictive analytics. A custom Neural Network architecture for time-series forecasting was developed to predict sales trends and stock requirements. Additionally, the project involved a RAG-based (Retrieval-Augmented Generation) assistant that provides personalized product recommendations by analyzing customer intent and purchase history. Enhanced customer engagement metrics and improved inventory turnover rates through more accurate demand forecasting and highly customized search experiences.

• Designed and deployed custom Neural Network architectures for high-accuracy time-series forecasting;
• Implemented RAG-based (Retrieval-Augmented Generation) architectures for personalized product discovery;
• Developed a backend API using FastAPI to serve real-time AI recommendations to thousands of users;
• Engineered complex prompt templates to maintain brand voice and safety across all customer-facing AI interactions;
• Built interactive internal dashboards using Streamlit to visualize model performance and forecast accuracy.

Standard hand-pose estimation models frequently fail in specialized industrial or AR/VR scenarios, especially when hands are wearing gloves or performing complex tasks. So the solution involved developing a massive synthetic dataset containing 1 million annotated images of glove-wearing hands. To bridge the "reality gap" between synthetic and real-world images, domain randomization and data synthesis techniques were used to simulate diverse backgrounds, lighting conditions, and textures. The model achieved significant performance gains on real-world images of glove-wearing hands, outperforming general-purpose models.

• Retrained and fine tuned State-of-the-Art Computer Vision models;
• Modified and optimized public hand-joint detection architectures for specialized industrial contexts;
• Developed custom loss functions to bridge the domain gap between synthetic and real-world image datasets;
• Authored and published research findings in the ECCV 2022 proceedings;
• Managed the end-to-end ML lifecycle from data collection and annotation to model deployment and benchmarking.