01The premise
Not prompts. Products.
The work starts with a job that needs doing, then adds models, tools, memory, schedules, evidence and approval boundaries only where they earn their place. A useful system should survive a second run and leave enough behind to understand the first.
02The systems
Agents are infrastructure with opinions.
Current work spans operational agents, private and local AI, multimodal retrieval, evaluation and repeatable media pipelines. Leonard for Euler Finance is the clearest public example: one named agent connecting market monitoring, code review, security checks, logs and recurring briefings.
03The lab
The hardware is part of the argument.
The home stack runs more than twenty-five services, including open models up to 80B, transcription, voice, image, video and multimodal RAG. It makes privacy, latency, cost and model choice observable instead of theoretical.