I'm building ARK — an open-source AI agent runtime in Go.
ARK solves three problems nobody else is solving together:
Context waste — AI agents dump 60K tokens of tool schemas into every prompt. ARK loads only what's relevant. 99% reduction.
Blind spending — companies know their total API bill but can't tell which agent decision costs what. ARK tracks cost per step and feeds it back into tool ranking.
One model for everything — every agent framework uses one model for the entire task. ARK routes each step to the right model: cheap models for tool calls, strong models for reasoning. Automatically.
Built from scratch in Go. 106 tests. 11 tools. 3 LLM providers. Verified cost tracking against OpenAI's billing dashboard. Single binary, zero dependencies.
Currently: CS @ DePaul University (AI concentration) | YC S26 applicant