real builds, real numbers.

Every serious AI workload was being billed per-token, with costs that scaled with usage and no ceiling in sight for a team that wanted to run agents continuously.

• Sized a 30B-class MoE model to a $500 Intel Arc Pro B60
• Tuned llama.cpp on Vulkan, picked quantization for the compute path, not the box-art VRAM number
• Benchmarked idle vs. loaded separately to get honest numbers

A production fleet generating real output at a fraction of API pricing, break-even against cloud costs in under two months of real use.

Every tool and agent on the fleet started cold — re-explaining context, re-discovering fixes that had already been worked out, with no shared source of truth.

• Built Neuralis: a persistent, shared memory every agent reads from and writes to
• Designed around supersede-never-delete and automatic write-back after every task
• Deduped before writing, so the brain doesn't drown in near-duplicate facts

Agents that start each task one step ahead instead of relearning what already cost real time to figure out — running live, past 130,000 memories.

Autonomous agents with tool access and no real guardrails are a liability wearing a demo — most "safety" in agent pitches is a suggestion the model can reason past.

• Clamped money and outbound network access by default, outside the model's own reasoning
• Built a recall → act → remember loop so failures don't repeat
• Put monitoring on a separate box with no stake in the outcome

An agent layer (Sentinel Prime) that fails safely instead of failing expensively — narrowing blast radius by design, not by hope.