Ketzai Agent: one system, different ways to plan

The challenge

I lead the team building Ketzai Agent. One of our biggest challenges was supporting different ways teachers plan—not just generating different documents.

Ketzai Teacher Lab helps teachers prepare lesson plans, didactic sequences, and worksheets. Ketzai has a presence in Panama and the Dominican Republic, with expansion to Colombia planned.

The differences go deeper than the final format. A planning approach can change the information we need from the teacher, the stages of an activity, the curriculum references, and the evidence used to assess learning. The system needs to preserve those differences while giving teachers a consistent way to work.

As Applied AI Lead, I lead the team’s work on the architecture, orchestration, and core runtime. I also work with the team on tools to test system behavior and evaluate its outputs.

A concrete example: several grades in one classroom

Multigrade planning is not the same lesson repeated for several grades. Students work at the same time around a shared activity, but each grade needs different tasks, support, and expected learning.

Our multigrade workflow represents that directly: each activity has a shared core and grade-specific instructions. It also accounts for different activity phases by education level and the teaching time available.

Other planning approaches need a different structure. Dominican primary planning, for example, brings together a thematic-unit matrix, confirmed curriculum references, a learning situation, teaching-strategy phases, and session timing.

These are not just wording changes. They affect the inputs, the generated plan, and the checks around it.

Share the foundation, keep the differences

One rigid workflow would make it hard to represent those differences. Completely separate systems would repeat the surrounding work.

Ketzai Agent uses a modular architecture: a shared foundation coordinates the request, while specialized workflows handle the planning details. LangGraph connects those workflows and maintains their state.

• Confirm the inputs. Generation receives a validated request, rather than having to infer the teacher’s choices again from the conversation.
• Use the right planning structure. Workflows and profiles provide the generation instructions, document structure, and checks needed for the selected context.
• Keep review with the plan. Specialized workflows handle their review and save steps, so different kinds of plans do not have to follow an identical process.

For Dominican primary plans, for example, code checks required context before calling the model and checks the approved document against confirmed information before saving. Prompts guide the educational content; code handles checks that should not depend on the model following an instruction.

What this makes possible

The shared foundation gives us a place to add planning capabilities without putting every detail into the central coordinator. The specialized workflows give us room to handle differences that matter in the classroom.

That is the point of the architecture: not making every teacher plan the same way, but supporting different approaches within one product. Teachers can review the material and request changes before saving it.

Explore my AI consulting and implementation services.

Tell me what you need →