Problem
Diagramming and literature review are the slowest, most error-prone parts of starting any research project. I wanted a single workspace where a student could brainstorm visually, watch the AI fill in adjacent concepts, and pull cited evidence (IEEE Xplore, Google Patents) into the same canvas without context-switching.
Approach
Three subsystems sharing one React Flow canvas. Diagram Bot grows the graph from any node using a structured "expand this concept" LLM prompt. Genie Assistant is a side-panel chat scoped to whatever node is currently selected. Research Assistant fetches and summarises IEEE Xplore / Google Patents hits and offers them as draggable nodes.
Implementation
Frontend in Next.js 15 with React Flow for graph state, server actions for orchestration, and Genkit for LLM tool-routing. I model every node as a typed record so the LLM tools read and produce strongly-shaped JSON. The hardest part was retrieval relevance — naive embeddings drowned in noise — so I added a re-rank step against the user-selected node label.
Results
Across pilot users, time to first usable concept map dropped from ~25 minutes to under 5. The "expand from this node" affordance is the most-used feature by 4×.
Lessons
Latency budgets matter more than capability in agent UIs — users will tolerate a worse answer if it streams in under 800 ms but abandon a perfect answer that takes 6 seconds. Building tool-routing on top of a typed schema (not free-form prompts) made the system 10× easier to debug.
Gallery
