DFFRNT AI Assistant: Air-Gapped Private AI Knowledge & Content Assistant

Meet the Team

The Client and Technical Advisor 

Client Name: Ben Dick

Senior Data Analytics Strategist at City of Ottawa

View Portfolio Website →

Technical Advisor: Arya Rahgozar

Adjunct Professor

The School of Engineering Design and Teaching Innovation, University of Ottawa

View Portfolio Website→

Prorotype Demo

Overview

Consulting firms like DFFRNT sit on years of institutional knowledge - proposals, resumes, case studies, project records scattered across disconnected storage with no way to search it centrally. Commercial AI tools like ChatGPT or Copilot could help, but DFFRNT's client-confidentiality obligations make sending that material to any cloud service off-limits. 

The challenge

How do you give consultants the experience of a modern AI assistant without letting any data leave hardware the client fully controls?

What We Built

The group (AirVaultAI) designed and delivered a fully air-gapped, private AI knowledge and content assistant that runs entirely on DFFRNT's own hardware, with zero outbound network connections during operation. 

The system:

• Searches the firm's entire document base using natural language and returns grounded, cited answers in seconds

• Drafts RFP-style resumes and proposal content grounded in real work history, not generic AI output

• Runs offline, end-to-end, on open-source components (Ollama-hosted LLMs, a Qdrant vector database, and a FastAPI backend)

• Requires no technical background to use - no command line, no cloud account, no configuration

How It Works

Documents (PDF, DOCX, PPTX, XLSX, CSV, TXT) are uploaded through a simple web interface, automatically split into searchable chunks, and embedded locally. 

When a consultant asks a question, the system retrieves the most relevant chunks by meaning (not just keyword matching) and generates an answer using a locally hosted language model - every answer traceable back to its source document and page.

[Insert 1–2 screenshots here: chat view with citations, and the document upload/tagging view]

From the Students

This course offered a valuable combination of conceptual grounding and applied experience for every member of our team. The lectures clarified how engineering design, requirements gathering, and validation connect into a single process, and they gave us a shared vocabulary for reasoning about problem statements, user needs, and iterative prototyping instead of moving straight into coding or implementation details. That framing mattered a great deal on a project like ours, where the hardest requirement wasn't a visible feature at all, but a guarantee that no data would ever leave the client's hardware that had to be argued for and proven, not just built.

Working on the Air-Gapped Private AI Knowledge & Content Assistant gave us hands-on practice with topics that are usually only discussed in theory, such as scoping a minimum viable product, designing for non-technical end users, and balancing new capabilities against the need for robustness, security, and testing. Building a fully offline RAG system for a confidentiality-bound client pushed us to translate an abstract privacy requirement into concrete, testable behaviour: zero network egress at inference, two-level document tagging with folder inheritance, mandatory source citations on every answer, and an explicit refusal behaviour when no supporting document exists. We experienced directly how client feedback can reshape both the interface and the underlying design: guidance from DFFRNT led us to replace zipped-file uploads with folder uploads, adopt a ChatGPT-like conversational interface instead of a bespoke one, and remove every trace of command-line interaction from the workflow. The project also forced us to coordinate frontend, backend, ingestion, and infrastructure work across six people, which strengthened our teamwork, communication, and version-control habits.

As a result, the course has advanced our academic and professional development in several ways. It reinforced the importance of clear documentation and explicit design reasoning, which will serve us well in future research and capstone work. It also gave us a concrete, real system we can describe in portfolios and interviews, demonstrating experience with Retrieval-Augmented Generation, vector databases, offline/air-gapped architecture, containerized deployment, and grounded content generation - all delivered in collaboration with an external industry partner. Overall, the combination of lectures and project work has made us more confident about taking on complex, multi-stakeholder software projects in our future studies and careers.