Product engineering for AI teams shipping beyond the demo.

The challenge was not just "displaying data," but defining how researchers interact with a new form of intelligence testing. I owned the end-to-end implementation of the visualization layer.

Product Impact:

• Visual Task Explorer: Engineered a complex SVG/Canvas based explorer allowing users to navigate hundreds of abstract reasoning puzzles.
• Evaluation Infrastructure: Built the "stadium" for model performance, creating high-density comparison matrices that organize results across dozens of models and the full evaluation dataset.
• Metadata Architecture: Designed the schema and frontend logic to organize disparate task metadata into a coherent browsing experience.

Researchers using Zotero had no way to leverage NotebookLM's AI features on their libraries. Since no public API exists for consumer NotebookLM, I reverse-engineered Google's internal Boq/WIZ RPC protocol — mapping undocumented batchexecute endpoints, session token extraction, and the Scotty resumable upload flow for binary files. I then engineered a full bridge system — a Zotero plugin, a Chrome extension, and a local IPC server — to enable seamless two-way sync between the two platforms.

Product Impact:

• Polling Bridge Architecture: Designed a reverse-RPC system where a Zotero-side HTTP server dispatches tasks to a Chrome content script running inside the authenticated NotebookLM session.
• Dual-Mode Strategy: Implemented both Consumer (Free/Pro) and Enterprise API strategies, auto-detecting the user's environment and routing tasks accordingly.
• Bidirectional Sync: Forward sync uploads PDFs, URLs, and metadata to NotebookLM; reverse sync imports AI-generated source guides and notes back into Zotero.