An AI Development Platform Utilizing HITL Verification: From Source Data to Tailored Domain Assistants

This paper introduces the Human-Enabled Learning & Evaluation Network for Assistants (HELENA), an end-to-end, locally deployable artificial intelligence (AI) development platform designed to transform proprietary or non-proprietary source data into specialized, tailored AI assistants. The HELENA framework integrates retrieval-augmented generation (RAG) to ground models and low-rank adaptation (LoRA) to instill response characteristics and formatting. The platform employs reinforcement supervised fine tuning (SFT) through a slow-burn human verification process that converts AI-generated synthetic data into verified training sets to ensure the highest level of accuracy and alignment. A critical advantage of the HELENA architecture is its portability; the system is designed to be deployed entirely behind secure firewalls, ensuring stringent data protection and eliminating external vendor vulnerabilities. Finally, this paper reviews the preliminary training progress of current assistants and details future capabilities, including knowledge graph integration, active knowledge elicitation, multi-agent collaboration, and the development of tool-using tailored assistants. Ultimately, HELENA serves as a secure and scalable blueprint for capturing institutional knowledge and deploying specialized AI.