In short
A research team needed AI across its meetings, chats and code, but its IP couldn’t be sent to an outside AI provider. We built Bastion, a private, on-premises AI knowledge assistant on company hardware, keeping $1M+ in IP private with $0 in external AI fees.
01Challenge: Knowledge that couldn’t leave the building
An established technology company’s specialized research team was rapidly expanding the intellectual property behind its competitive edge. Much of its newest knowledge lived in meetings and group chats rather than formal documents. Leaders manually reconstructed progress across disconnected systems, while important findings moved slowly between research efforts and up to the C-suite.
AI could accelerate the team, but the information needed to make it useful was also the information the company could least afford to expose. Conversations, code, and intellectual property had to remain on employee devices or company premises.
The company wanted to use AI to compound its information advantage without giving that information advantage away.
02Solution: Private, on-premises AI
789 Labs built Bastion, an always-available knowledge assistant powered by on-premises instances of DeepSeek-V4-Flash. The models ran on company-owned consumer hardware, keeping sensitive information inside the company’s boundary and eliminating external AI usage fees.
Custom integrations connected the company’s self-hosted communication, knowledge, and code systems. Near-real-time ingestion kept Bastion current as conversations, documents, and engineering work changed. Permission-aware retrieval ensured that researchers received answers based only on information they were authorized to access.
An agentic layer coordinated data, interpreted requests, retrieved evidence, and executed recurring tasks. It handled the operational challenges of running AI privately, including tuning the local AI engine for the company’s hardware and balancing limited computing capacity between processing new company knowledge and answering concurrent requests from the team’s chat software. Responses cited the documents, messages, and code behind them, giving researchers and executives an evidence-backed view of the work.
03Impact
Bastion turned fragmented research activity into shared, current knowledge. Researchers could build on parallel efforts, while the C-suite gained a more complete view for high-impact decisions. The company achieved this without sending its intellectual property to an external provider.
- $1M+ in company IP kept private from Anthropic and OpenAI
- 0 company data sent to external AI providers
- $0 paid in external AI usage fees