Anton Fedoruk

Hi! I'm an Engineering Lead at kay.ai, building pragmatic systems and doing applied AI research.

Previously, I was a founder at Array AI, working on learning algorithms inspired by the human brain 🧠. Earlier, I founded WarSpark Studios and Blaster Studio, shipped mobile games reaching 7M+ players, and was featured by the App Store (Game of the Day in the U.S.).

Research

Code Crew

A famous-programmer code review crew for Claude Code, Codex, Cursor, and Hermes. Blind Knuth + Hickey + Torvalds passes with a diff-grounded verifier beat single-pass review recall by +7pp (paired n=50, p=0.004).

ReleasedGitHub

ROIBench

Autonomous agents iterate against three metrics (Activation ROI, Payoff, Retention Signal) until the app delivers real, measured value. Loss functions for products.

In progressroibench.com

Guaranteed Long-List Extraction via Item-Count Adaptive Chunking

Five-stage pipeline avoiding truncation via adaptive chunking. 100% recall on FCC Invoices and 10,000-claim synthetic tables.

In preparation

LongListBench: A Benchmark for Long-List Entity Extraction Under Layout and OCR Noise

Anton Fedoruk, Serhii Shchoholiev, Akhil Mehta

80-document benchmark for long-list extraction from semi-structured PDFs. Includes ground-truth JSON, OCR transcripts, and 7 failure modes.

Internal reviewPDF GitHub

Story

AI has been my childhood dream. In 2020 I committed to it full-time after years of building game studios as a way to learn, ship, and fund long-term research.

I started with visual understanding and kept returning to a hypothesis: saccade-like attention is key for making vision systems efficient and human-like.

During COVID, I was doing AI research for my startup Array AI in Kyiv. When the war in Ukraine started, everything changed quickly. In the following year, I joined kay.ai as a contractor and was soon invited to the U.S. to become the startup's first official employee.

Outside of work, I like running and pair-vibe-coding with my son.