fumpa7190/5day-ai-agents-capstone-project_v2

AI-powered offline classroom resource generator for rural teachers in Papua New Guinea using lightweight local LLMs and a multi-agent workflow # PNG Classroom Resource Generator An AI assistant that helps teachers in Papua New Guinea schools generate lesson plans, activities, teacher notes, and assessments - grounded in the real PNG Standards-Based Curriculum, and running entirely on a light local LLM (`google/gemma-4-e2b` via LM Studio) so it works without a paid cloud API or reliable internet access. The agent layer is built around **ADK Agent Skills** (`load_skill_from_dir` / `SkillToolset`): each agent's procedural knowledge - the rules and output structure for the resource type it generates - lives in a `skills/*/SKILL.md` file it loads at runtime, rather than being inlined into a large prompt. The implementation follows `specs/png-classroom-agent-workflow.md`, the source of truth for the agent workflow, guardrails, and Definition of Done. For house rules and the Skills architecture, see `AGENTS.md`. ## How it works ``` Teacher Selection | Curriculum Matching - deterministic, no LLM (services/curriculum_store.py) | Lesson Planning Agent - LLM [orchestrator/pipeline.py | run_lesson_plan_stage()] v (teacher clicks "Next") Activities Agent -> Teacher Notes Agent - LLM, sequential [run_activities_ | notes_stage()] v (teacher clicks "Next") Assessment Agent - LLM [run_assessment_stage()] | v Review and Alignment - deterministic, no LLM (services/review.py) | Combined Lesson Pack -> rendered, downloadable (Word or PDF) ``` Curriculum Matching and Review are plain Python, not LLM calls - so curriculum facts can never be invented and alignment checking is identical every run. Each of the four `LlmAgent`s carries a `SkillToolset` scoped to exactly two skills: a shared `png-classroom-conventions` skill (resource-availability defaults, PNG context, currency, plain-text math, mixed-ability, language level) and one task-specific skill (`l …