kusp is a personal research project focused on learning agentic engineering through building. 

Past projects include a 9-agent system on OpenClaw, a 5-agent system on Letta, and a fully custom 5-agent system built from scratch with no framework. Along the way, an experiment on adaptive retrieval optimization produced a novel finding: Thompson Sampling can learn retrieval weight preferences from LLM self-assessment rather than human feedback. That experiment ran across 3 versions and 2,200+ episodes, resulting in an open-source codebase and a research paper.

The project produced a controlled experiment (2,200+ episodes across 4 conditions), an open-source codebase, and a research paper. All work is conducted on personal time, using personal hardware and publicly available data, with open-source tools or personal subscriptions.

This page is a repository for sharing learnings, experiment updates, and insights from building agentic systems from scratch on a budget ($50/month in API costs, consumer hardware).

Website: kusp.dev
Code: github.com/kusp-dev
Writing: substack.com/@alfonsodirocco

kusp

Org chart

Teams

Offices

kusp

Org chart

Teams

Offices

Related companies