Ivan Leo

Ivan Leo is a developer-experience engineer at Google DeepMind, where he helps developers build with Gemini and evaluates autonomous agents. Previously a core maintainer of Instructor and an engineer at Manus, he builds the interfaces, evaluations, and orchestration systems that make language-model applications reliable.

Leo began his career as a front-end engineer at Credit Suisse in Singapore. He taught himself machine learning through online courses and an AI paper-reading group that met at 3 a.m. locally, eventually helping organize an Asia-friendly edition. While still working in banking, he contributed documentation and articles to Instructor, an open-source library for producing validated language-model outputs with Pydantic. Those contributions led to a full-time role working with Jason Liu and maintaining the project, a transition he describes in his account of breaking into AI.

Leo also contributed to Kura, which analyzes language-model conversations by labeling user behavior and clustering embeddings while limiting exposure of personally identifying information. At Manus, he later applied related techniques to multilingual customer-support failures.

After joining Manus in July 2025, Leo worked on Mail Manus, file suggestions, the Manus API, and Stripe integrations for generated websites. He became a principal developer on the API after identifying and fixing attachment and connector issues, learning production Go as his responsibilities expanded. His reflections on building at Manus emphasize shipping working prototypes, testing integrations with actual transactions, and owning documentation and support alongside implementation.

His AI Engineer Code workshop showed how those principles translate into operational agent infrastructure: asynchronous tasks, durable identifiers, file uploads, and webhooks. He built a Slack integration that linked conversation threads to persistent agent sessions and returned generated files to the correct thread, then connected receipt extraction to company expense policies stored in Notion. Other demonstrations included a French-practice application and a personalized conference planner combining browser automation, semantic recommendations, and calendar integration. He emphasized explicit authorization for browser access and careful handling of sensitive files.

Leo subsequently joined Google DeepMind’s developer-experience organization, focusing on Gemini and autonomous-agent evaluation. He has also helped build 65Labs, a Singapore community organizing AI hackathons and developer gatherings.

Ideas that shape his work

• Structured outputs as software contracts. Leo explains how Instructor converts Pydantic models into provider-compatible schemas, making probabilistic model responses easier to validate and integrate into conventional applications.

• Start with simple, measurable evaluations. His approach to lightweight evaluation begins with direct checks: valid parsing, required fields, relevant retrieval, and correct tool selection.

• Webhook-driven agent orchestration. Polling helps teams prototype quickly; persistent task identities, callbacks, and thread-aware state become essential as concurrent workflows scale.

• Benchmark retrieval before adding complexity. In a comparison of terminal search and SQLite full-text search, his terminal-based approach cost less and required fewer conversation steps than a naive SQLite baseline across approximately 300,000 tokens of developer documentation.