When we started building Celestix AI, the obvious choice was cloud-based AI. OpenAI's GPT-4 was the clear market leader. Cloud deployment would be simpler. Scaling would be easier. But we chose a harder path — and here's why.
Your Bid Data Is Your Competitive Advantage
Think about what goes into a bid price: your labor rates, your overhead calculations, your profit margins, your subcontractor relationships, your risk assessments. Now imagine uploading all of that to a cloud AI service that also processes your competitors' data.
Even with enterprise-grade security, the risk calculus doesn't work. Cloud AI providers train on user data (or could change their policies to do so). Your pricing strategies, cost structures, and competitive intelligence would exist on servers you don't control. In federal contracting, where competitors bid on the same solicitations, this is unacceptable.
The Ollama Revolution
Ollama changed the equation. For the first time, we could run production-quality language models locally — on hardware that contractors already own. Our system runs on Llama 3.1 8B, which delivers strong reasoning capabilities while fitting comfortably on a machine with 16GB RAM.
The performance trade-off is real but manageable. A full 4-round analysis takes about 40 minutes on a modern laptop with local AI, compared to perhaps 10 minutes with cloud GPT-4. But that extra 30 minutes buys you something priceless: absolute certainty that your data never left your machine.
Zero Network Dependency
Local AI means no internet required for analysis. Your Celestix system works on an air-gapped network, in a SCIF, on a plane, or anywhere else. This also means no API rate limits, no outages because a cloud provider is having a bad day, and no surprise cost increases when your AI vendor raises prices.
How We Optimize for Local
Running 32 AI agents locally requires careful engineering. Each agent doesn't run a full LLM instance — that would require 32× the memory. Instead, we use a shared inference engine with agent-specific prompts, knowledge bases, and reasoning frameworks. The LLM provides the reasoning backbone; the agent architecture provides the specialization.
We also aggressively cache and pre-compute. Common calculations (prevailing wage lookups, RSMeans data, regional factors) are stored locally in SQLite. The LLM is only invoked for judgment calls — not for data retrieval.
The Privacy Guarantee
Our privacy model is simple: your data never leaves your machine. Not for analytics, not for model improvement, not for anything. We don't have servers to breach because we don't have servers. Your Celestix installation is a self-contained intelligence system.
For federal contractors handling CUI (Controlled Unclassified Information), this architecture meets NIST 800-171 data handling requirements by default — because there's no data transmission to secure.