Indonesian students continue to perform below the OECD average in scientific literacy and problem-solving on PISA 2022, and existing classroom interventions rarely combine STEM integration with Artificial Intelligence (AI) scaffolds within a single coherent design. This study developed, validated, and tested the effectiveness of an AI-assisted STEM learning tool (STEMI-AI) for improving the scientific literacy and problem-solving skills of seventh-grade students at a junior secondary madrasah in Kudus Regency, Central Java. The Plomp development model was applied across three iterative phases: preliminary research, prototyping, and assessment. Participants were 32 students of class VIIA (experimental group) and 32 students from a parallel class (control group), purposively sampled. Data were collected through expert validation sheets, teacher and student practicality questionnaires, a scientific literacy test adapted from the PISA framework, and a problem-solving test based on Polya's four-step rubric. Data were analyzed using V Aiken for content validity, percentage analysis for practicality, N-gain and analysis of covariance for effectiveness, with Cohen's d as the effect-size index. Expert validation produced a mean V Aiken of 0.91 (highly valid), practicality reached 88 percent (teachers) and 84 percent (students), and the experimental class showed significantly greater gains than the control class on both scientific literacy (N-gain = 0.58, medium) and problem-solving (N-gain = 0.54, medium), with Cohen's d of 1.12 and 1.05 respectively. The integrated AI-STEM design is feasible and effective for junior secondary madrasah classrooms when AI components are embedded as pedagogical scaffolds rather than as standalone aids.