Pavel Kazantsev is an AI and machine-learning engineer who publishes evaluation systems and research notes. His PromiseDiff project checks whether digital-product files support the claims made in a sales listing.[1]
Start with PromiseDiff – audit listings against delivered files →
GitHub 
X Contributions
PromiseDiff accepts a listing and a PDF or ZIP download, then checks claims against the supplied files. Direct inspection can establish facts such as page counts and document dimensions, while a local language model helps interpret semantic claims. Results distinguish supported, contradicted and unverified assertions. Missing extracted text or an ambiguous file does not automatically become evidence that a seller’s claim is false.[3]
Ideas
Kazantsev evaluates the local application on ten synthetic cases and retains raw responses and timings. He reports nine matching labels, including a paraphrase the system left unverified. The small test set is a reproducible scaffold, with limited evidence about real-product accuracy. His broader portfolio similarly emphasizes inspectable systems, public limitations and results that did not survive scrutiny.[3][1]
Tools, writing & talks
Writing, research & talks
Start here
PromiseDiff – audit listings against delivered filespkazantsev.comA case study of file inspection, local-model interpretation and evidence-backed claim status.
Read the PromiseDiff case study and compare measured facts with the unverified category before trying it. Running the local app requires the documented backend and local-model setup; its ten synthetic tests are a scaffold, not a real-world accuracy guarantee.
Author and project creator[3]Original source