Arjun Singh

Arjun Singh is co-founder and chief executive of Superconductor, which builds collaborative coding-agent software that lets engineers and their colleagues share control of automated development. He previously co-founded Gradescope, the AI-assisted grading company acquired by Turnitin in 2018, and later served as Turnitin’s chief strategy officer.

Singh earned a computer science doctorate at the University of California, Berkeley, where he researched robotics and taught artificial intelligence. While teaching, he worked with professor Pieter Abbeel and fellow teaching assistant Ibrahim Awwal on Pandagrader, an early tool for reducing the administrative burden of grading. In 2014, he met Sergey Karayev while they were teaching an AI graduate course; they subsequently developed the project into Gradescope.

By the time Turnitin acquired Gradescope, the platform had processed more than 60 million answers at more than 600 institutions. Singh later reunited with Karayev and former Gradescope colleagues to create Superconductor, applying their experience building collaborative products to agent-assisted software development.

• Multiplayer agentic engineering: Singh treats coding agents as shared team infrastructure. Persistent sessions move across Slack, desktop and mobile applications, and GitHub without losing context, allowing engineers, support staff, and growth teams to inspect artifacts, contribute feedback, and track who has reviewed proposed changes.

• Turning customer conversations into prototypes: Meeting bots can identify product ideas, connect them with existing work, and generate prototypes or pull requests. Singh demonstrated how feedback about clearer evaluation criteria became a proposed change to a ticket form, giving the team concrete software to assess before deciding whether to ship it.

• Isolated cloud sandboxes: Agents operate independently of individual laptops with narrowly scoped credentials, controlled network access, and human approval for unfamiliar destinations. These boundaries also let nontechnical colleagues initiate changes without installing development environments or exposing sensitive systems.

• Codebase-specific agent evaluation: Singh benchmarks competing models and agent harnesses against representative pull requests from his team’s own repository, comparing quality, cost, and execution time. A Python-heavy public benchmark may say little about a Ruby on Rails application; project-specific testing supports better task routing and reduces dependence on any single provider.

His approach to collaborative coding agents keeps automated work visible across the organization while requiring human review before agent-generated code ships.