Ihza MahendraResume
Back to Agentic & multimodal systems
Agentic & multimodal systems

Just-in-time learning break inside a work task

A support flow a person can step into mid-task when they're stuck: the same AI teammate they've been working with pauses the task and walks them through exactly what they're missing before they go back to it.

What it is

A learning flow a person can step into when they get stuck partway through a work task. The same AI manager they've been working with (see the multi-character AI team project) pulls them off the task for a moment and walks them through the specific concepts they need in order to finish it. It's structured: each concept the task depends on has a short pre-made video, and the videos play in the order that builds up to finishing the task. After the videos, the manager offers a live practice session where the person can talk through what they just watched with an AI coach before returning to the task.

What it's for

Pushing through a task you don't understand teaches the wrong lesson: you either guess your way to a pass, or guess your way to a fail. A person needs a way to step out of a task, get the concepts they're actually missing in the right order, and come back without losing their place. This flow is designed as a break from the task, not a separate program: same AI teammate, same memory of the conversation, just a different goal for a few minutes, teach the concept, instead of pushing through the piece of work.

How it was built

Triggers when the person explicitly asks their AI team for help, or when the system detects they need it and routes them in automatically. The manager pulls the list of concepts the task was built to teach, picks the right pre-made video for the first gap, and plays it. When the video ends, the next one queues up in the order the task needs them. After the sequence, the manager offers a live practice session so the person can talk through what they saw with an AI coach beside them. The flow runs inside the same conversation system as the main task (see the multi-character AI team project), sharing the same character, memory, and profile, so the manager doesn't feel like a different person when the flow switches. Time-based check-ins (after 20, 30, and 45 minutes) make sure the session doesn't drift. Everything that happens still feeds back into how the person is read, including how much they leaned on this flow for the task, which factors into a read on how independently they can work.

My role

Co-contributor on the routing into this flow from the main task, the video-sequencing logic, and the wiring back into how the person is read when the session ends.

Built with
PythonFastAPIWebSocketsGeminiTemporalPre-generated video sequencing

Want the full technical depth, the tradeoffs, what broke, what I'd do differently? Ask the agent about this project.