Greg Kamradt

Kamradt studied business at Loyola Marymount University and later took a data-science bootcamp in 2015. A 2016 PyData biography places him on Salesforce’s product data-science team as a senior growth analyst. He describes subsequently becoming the first business and operations hire at the financial-software company Digits, with work spanning analytics, monetization and product decisions. That background helps explain his emphasis on testing what AI can do in a concrete workflow.[5][6]

His personal portfolio includes InterAuth, which gives an AI agent scoped read access to a Google Doc, and Agent Traffic Control, a view of agents doing work in parallel. These projects connect his evaluation work to practical agent workflows.[1]

Measuring adaptability

Kamradt co-authored the ARC Prize 2025 technical report. ARC evaluations test how systems adapt to new tasks, making the benchmark useful alongside tests of memorized knowledge. The report documents the competition and its results.[3]

Testing whether a long context is usable

His Needle in a Haystack benchmark inserts a known fact into a longer body of text and asks a language model to recover it. By varying both the total length and where the fact appears, a reader can distinguish a model’s advertised input capacity from reliable retrieval across that input. The repository now supports single facts, multiple facts and chains of identifiers. Those are controlled retrieval tasks, so success should not be read as proof that a model understands every kind of long document.[7][8]

Kamradt’s project portfolio also explores how agents receive information. InterAuth provides a disposable, read-only Google Doc share address, while agnts.sh serves a human page, Markdown or JSON according to the client’s request. Both make a specific interaction easier: giving software the source material it needs without asking it to reconstruct a conventional website first.[7]