Background Laryngomalacia is the commonest cause of stridor in infants. Classification systems are based on assessment of laryngeal subsites responsible for supraglottic collapse. They suffer from variable inter-rater reliability and correlate poorly with clinical severity. Artificial intelligence outperforms humans in many diagnostic settings. In this pilot, we develop a convolutional neural network to classify laryngomalacia. Methods 100 de-identified recordings of laryngoscopies in infants were retrieved and mirrored to create 100 duplicates. Two paediatric otolaryngologists graded severity, epiglottis shape/position/score, arytenoid score, aryepiglottic folds, supra-arytenoid mucosa and tongue-base. We assessed intra- and inter-rater agreement using Cohen’s kappa. Successive frames were extracted in greyscale 100x100 pixels. A CNN was created using TensorFlow2.0 with rectified linear activation and four layers of convolutional 2D-neurons. 80% of recordings were used for training and 20% for validation. Results Intra-rater agreement (κ=0.78) on overall severity was substantial. Intra-rater agreement was at least substantial on all anatomical sites (κ:0.72-0.90) except aryepiglottic folds (κ=0.05). Inter-rater agreement on severity was fair (κ=0.35). Training data for tongue-base (κ<0) and epiglottis position (κ=0.20) were excluded due to poor agreement. There was otherwise fair to substantial agreement (κ:0.33-0.68). After training, the model accurately assessed laryngomalacia severity in 83.2% of test cases. Model accuracy for anatomical subsites varied from 67.6% - 84.9%. Conclusion This pilot study demonstrated a CNN that can accurately assess laryngomalacia severity and grade some anatomical subsites, with limited training. A larger dataset and high-fidelity images might improve performance. Future iterations could combine endoscopic and clinical data to predict clinical severity and likelihood of surgery without reference to classification systems designed for human reviewers.