Wedge
A task and habit product built around retrieval-augmented context and a small robot assistant.
Wedge tested a more context-aware productivity companion: one that could turn a task into smaller next steps instead of merely storing a list.
The situation.
Generic task systems can preserve work without helping a person decide how to begin. Wedge explored whether retrieval-augmented context and a small robot assistant could make task suggestions, habit management, and productivity questions feel more responsive.
I built the prototype in May and June 2024 as an AI-powered task manager that uses retrieval-augmented generation to turn larger tasks into suggested subtasks.
What I did.
I developed the prototype and described it publicly as a productivity platform centred on Wedge, a robot that helps with task suggestions, habits, and productivity questions.
It remained a personal prototype, giving me space to explore how stored context, suggested next steps, and a visible assistant character could work together.
How the work unfolded.
Focused on the next action
Explored subtask generation so an ambiguous task could become a more approachable sequence of steps.
Used contextual retrieval
Applied RAG as a way to ground task suggestions in stored context rather than treating every request as isolated.
Gave the assistant a product role
Used a robot persona to make the assistant a visible part of the interaction model, not simply an embedded text box.
What happened.
By June 2024, Wedge had become a working prototype using React, FastAPI, Azure Cosmos DB with MongoDB, Azure OpenAI, and LangChain.
It was an early technical and product sketch of contextual productivity assistance, especially the idea that an assistant should help turn an unclear task into a useful next action.
What I learned.
Context only helps when it changes the next step
RAG is not a product outcome by itself; the test is whether it produces a more useful action for the person.
A companion needs boundaries
Giving an assistant a friendly role should not obscure what information it can actually use or how reliable its suggestions are.
Related links.
Repositories, demos, event pages, posts, and other places connected to this project.