2026 ASEE Annual Conference & Exposition

Generative AI is increasingly used as an assistant for knowledge work, supporting brainstorming, drafting, revising, coding, and transforming ideas into usable artifacts. In engineering education, these capabilities create an opportunity to reduce faculty time spent producing instructional materials. At the same time, responsible use must satisfy competing constraints: maintaining instructional quality, avoiding hallucinated or biased content, supporting ethical AI practices that students can observe (e.g., traceable sourcing and explicit review), preserving faculty pedagogical judgment, and operating within real world limits of time, cost, and privacy requirements.

This paper reports the implementation of POSED (Plan, Outline, Sources, Evaluation, Draft) system, an agentic, human-in-the-loop (HITL) curriculum-development workflow realized as a reproducible low-code system in n8n platform. POSED uses visual orchestration and deterministic control frames around probabilistic AI generation calls, with pause-and-approve checkpoints that enforce faculty-led governance and preserve editable intermediate artifacts. Rather than treating generative AI as an autonomous content producer, POSED positions it as a constrained collaborator embedded in a faculty-approved, process-guided workflow.

POSED encodes multi-stage prompting, artifact handoffs, and governance gates into an executable pipeline so instructors do not need to manually coordinate prompts, intermediate files, and model configurations across tools. A central systems contribution is treating model choice and settings as first-class workflow decisions: POSED supports modular routing across heterogeneous models and tools to balance capability, time efficiency, privacy/IP constraints, and operational (token/API) cost. In the evaluation runs reported here, most generation used cloud LLM APIs, while workflow execution (n8n) and retrieval storage (local Qdrant) were self-hosted; optional local-model execution is supported for privacy-sensitive steps.

This design also addresses a reliability tension rooted in LLM sampling: more exploratory generation can improve idea exploration and fluent drafting while increasing the risk of unverified details; more conservative settings can reduce unsupported elaboration but still do not guarantee correctness. Accordingly, POSED treats reliability as a workflow property enforced by deterministic control logic around probabilistic generation: retrieval-augmented generation (RAG) is grounded in instructor-approved sources, structured critique is applied before final drafting, and HITL approval gates ensure scope, structure, sourcing, and final outputs remain faculty-controlled and traceable. POSED also preserves teaching style through a versioned professor-persona artifact that can accumulate refinements across runs, reducing homogenization without shifting the instructor into post-hoc “editor only” labor.

We implemented POSED as an n8n-based workflow with more than 200 nodes and evaluated it by generating lecture slides and instructor notes across course contexts in Electrical and Computer Engineering and Computer Science. Within this scope, observed faculty labor for this instructor decreased from an estimated 10-20 hours per lecture for manual preparation of new or substantially revised materials to approximately 2 hours of active review and edits, with marginal computational AI cost under \$1.5 per lecture.

We further report a horizontal comparison against three realistic direct-generation baselines (ChatGPT, Gemini Gems, and NotebookLM) and a vertical comparison across Aug.~to Dec.~2025 that illustrates capability drift and motivates whyworkflow-level governance remains necessary even as models improve.