Ido Pesok is an engineer at Cognition working on autonomous software verification for Devin, its AI software engineer. Previously an engineer on Vercel’s v0, he builds systems that evaluate AI-generated code and enable coding agents to prove their changes actually work.
From developer tools to AI alignment
Pesok studied computer science at California Polytechnic State University, San Luis Obispo, graduating in 2024. His early engineering experience included frontend development at Level 2 Designs, backend work at Amazon, and document search, summarization, and medical software at Sequoia Port. At Cal Poly, he also helped build Enroller, an alternative course-enrollment application with AI-powered access to university course information.
In 2023, he co-founded Energize AI and collaborated with Ethan Shaotran and Sam Jones on Aligned, a project for collecting public input about how AI models should behave. Included in OpenAI’s democratic-inputs initiative, Aligned explored large-scale participation and Community Notes-inspired model guidelines.
He also created ZSA, an open-source library for type-safe Next.js server actions that combines schema validation, middleware, authorization, error handling, retries, and React Query integrations.
At Vercel, Pesok co-authored an influential framework for eval-driven development, combining deterministic checks, human assessment, and model-assisted grading to improve probabilistic applications. He also helped explain v0’s composite model architecture, which supplements frontier models with retrieval, project context, specialized editing, and automated error correction.
By early 2026, he had joined Cognition, where he develops verification systems for autonomous coding agents: software that runs applications in cloud environments, tests relevant workflows, and returns inspectable evidence that changes behave correctly.
How Pesok approaches AI reliability
Evaluate the product users actually encounter. Pesok builds evaluation datasets from production logs, feedback, support channels, and community reports, distinguishing meaningful edge cases from requests outside an application’s purpose. His AI Engineer World’s Fair session demonstrated how unexpected combinations of otherwise ordinary inputs can break a convincing AI product demo.
Keep evaluations close to production. Stable user examples should remain separate from changing prompts, retrieval strategies, and models. Shared middleware makes evaluation runs representative; deterministic pass-fail scoring simplifies debugging; continuous-integration checks expose improvements and regressions before changes merge.
Make coding agents substantiate their work. At Cognition, Pesok uses source-grounded test plans to identify affected functionality before agents exercise an application. Agents record expected outcomes, classify assertions as passed, failed, or untested, and return annotated screenshots or recordings. Deterministic scripts handle repetitive setup such as authentication, while execution guardrails prevent fabricated paths, irrelevant exploration, misleading screenshots, and browser interactions that bypass the actual user experience.
A fruit-counting app shows why successful demos are not enough—and how representative data, shared production logic, simple scoring, and regular evaluation make improvements measurable.
