Team B3 - Skills 3
A team-built assistant that helps technical sales teams review incoming emails, retrieve grounded context, and draft accurate customer responses.
This was a team project for Thomas More University of Applied Sciences. We built an AI-powered technical sales assistant that processes customer emails, classifies them by urgency and status, retrieves relevant company knowledge, and helps generate a draft response that can be reviewed, edited, and sent by a user.
My responsibility was the AI development part of the project. I was in charge of most of the AI training work, the grounding of the model with retrieval-augmented generation, and the implementation that connected the AI layer with the backend workflow.
AI-generated email drafts for technical sales and support questions.
RAG grounding with Redis vector search and sentence-transformer embeddings.
Interactive chat for follow-up questions, feedback, and additional context.
Verified Q/A flow that stores approved responses for future retrieval.
Prepared the model to answer from trusted project and company context using RAG, Redis vectors, document embeddings, and quality gates.
Worked on most of the training side, including proof-of-concept LoRA training, dataset preparation from PDF and email content, and testing model adaptation paths.
Connected the AI layer to the FastAPI backend through an MCP-based flow, allowing the application to request grounded drafts and store approved responses.
Documented what worked, what was blocked, and what future teams should improve, including hybrid search, customer context, and production training integration.
Team AI application
Zyntek
AI Development, Training, and Backend Integration
Skills 3
Thomas More University of Applied Sciences
The AI needed to answer with reliable company-specific context rather than generic model knowledge.
I grounded the model using a RAG setup with Redis vector search and local embeddings, then connected it into the backend draft-generation flow.
Training and deployment had compatibility limits between LoRA adapters, Ollama, and vision models.
We proved the training path, documented the integration blockers, and separated text and vision model paths so the production system could keep moving.