A PRIVACY PRESERVING PERSONAL AI ASSISTANT AND THE COST OF CLOUD MODEL INFERENCE: ARCHITECTURE, CAPABILITY TRADEOFFS, AND EVALUATION PROTOCOL

Personal AI assistants need reliable research and controlled access to private data. This report specifies a local-first Windows architecture and examines the capability, privacy, and operating-cost tradeoff between local Llama 3.2 3B through Ollama and self-hosted Kimi K2.5 on rented NVIDIA GPUs. We analyzed a supplied project design summary, mapped trust boundaries and failure modes, and calculated illustrative costs from a documented eight-H200 deployment and published rental rates. The design separates public research from authenticated browser sessions, enforces permissions outside the model, and requires evidence-supported answers. At an illustrative $4.59 per GPU-hour, eight H200s cost $36.72 per running hour, $4,406.40 for four hours daily over 30 days, or $26,438.40 for continuous 30-day operation, before storage and operations. An assumed 100 W incremental local draw for four hours daily at $0.15/kWh costs $1.80 monthly on an existing PC. These are scenarios, not measured deployments. Kimi K2.5 offers multimodal and agentic capabilities that a text-only 3B baseline should not be assumed to match; no matched end-to-end test was performed. Stronger cloud inference can impose substantial dedicated GPU rental cost and transfers prompts to a cloud operator. Local inference preserves a narrower data boundary at far lower incremental cash cost but requires direct task-quality evaluation.