Abhi Arya is a software engineer and the co-founder and former CTO of Opennote, the AI learning startup acquired by Reducto in May 2026. His work spans infrastructure and interfaces: spacewalk guidance, browser automation, collaborative learning software, and document-processing agents. At AI Engineer World’s Fair 2026, he presented work on product at Reducto, explaining how agent tools can help people understand and guide automated work.
From spacewalk guidance to browser automation
At the University of California, Irvine, Arya studied computer science and engineering. He was the first-listed author of Polaris, a collaborative research project developed for NASA’s 2024 Student User Interface Technologies for Students challenge. The proposed system paired a head-mounted display for astronauts with a local mission-control console for their support team. Navigation and operational guidance depended on both interfaces, making human coordination part of the design. Polaris was a student research system, not deployed astronaut equipment.
His software engineering internship at NASA Johnson Space Center, from August through November 2024, covered autonomous-control algorithms, software verification, and tools for working with downlinked Gateway space-station data. His responsibilities ranged from embedded computing constraints to React and TypeScript interfaces for human operators.
In spring 2025, Arya interned at Browserbase as an early engineer on Director, its browser-automation product. He worked on browser infrastructure, session management, and the code-generation pipeline—the machinery connecting an automation request to software that could act on the web.
From collaborative learning to document infrastructure
Arya co-founded Opennote with Rishi Srihari and Vedant Vyas. The startup grew out of their student experience and joined Y Combinator’s Summer 2025 batch. Its AI tutor, Feynman, operated inside a student’s notes, using questions, feedback, and visual explanations to help develop understanding. Collaborative journals let students work through material together.
As CTO, Arya worked across backend infrastructure, real-time collaboration, agent systems, developer interfaces, and a text-to-video generation pipeline. Those systems connected the tutor to a learner’s material and ongoing work, including work shared with other students.
The founders’ May 2026 acquisition announcement explained how helping students make sense of notes had led them toward a broader problem: making large amounts of scattered information usable. Joining Reducto carried that work into document infrastructure for people and agents. The announcement also outlined the retirement of Opennote’s standalone learning product as the team joined Reducto.
Giving people visibility into automated work
Two projects illustrate Arya’s approach to software that performs complex tasks on a user’s behalf:
Natural-language model deployment: Arya co-created Runway with Srihari at TreeHacks 2025. The prototype connected dataset generation, model selection, training, and deployment through natural-language requests. Arya’s individual contribution included the real-time logging and search system, giving users visibility into a process spanning several technical stages.
Agent experience and responsibility: Arya’s account of rebuilding Reducto’s MCP integration centers on who bears the cost when an agent makes a mistake. The first server exposed a tool for every API endpoint. It demonstrated well, but team use produced confidently assembled workflows that people could not explain. Arya described deleting that version and rebuilding it with fewer tools representing meaningful work, plus a snapshot tool exposing live state. Visible confidence scores and requests for clarification helped users recognize uncertainty and supply missing information. Instrumentation then fed usage back into improvements; he reported that the sales team adopted the system without being asked.
Across these projects, Arya works on both the infrastructure that performs a task and the interfaces that let people follow its progress, recognize uncertainty, and intervene.
Reducto’s first MCP server could build document workflows confidently, but nobody could explain them. The rebuild moved product knowledge into tools, exposed uncertainty and used real sessions to improve future work.
A tool for every endpoint gives an agent access, but can leave workflow meaning and assembly entirely to the model. Encode known product patterns in tools that users can understand.
Pair structured operations with live-state snapshots so the agent can reference the actual pipeline, including specific errors, instead of guessing its current configuration.
Handle output uncertainty with extraction feedback and task uncertainty with clarifying questions. Evaluations should penalize unsupported assumptions rather than reward confident completion.
Logged overrides, low confidence and questions can feed longer-term guidance. The proposed learning loop improves the context around future agent work.
