Design ModernBERT-Pro, an enhanced encoder-only transformer incorporating Multi-head Latent Attention (MLA) for global layers, fine-grained Mixture-of-Experts FFN layers, Mamba-2 hybrid blocks for linear-complexity local mixing, and a multi-stage curriculum pretraining pipeline. The target is a ~350M-param model (with ~2.5x effective FLOPs via spar

Design 'EfficientDenseNet' — a DenseNet variant under 10M params that incorporates VoVNet-style one-shot aggregation (OSA), depthwise separable bottlenecks, squeeze-and-excitation attention, LayerNorm pre-activation, and aggressive compound scaling. The goal is to train 2-3× faster than DenseNet-121 while maintaining or improving accuracy on ImageN

Design a hybrid ML architecture that uses a GBDT ensemble as a hard-routed geometry extractor (producing one-hot leaf embeddings for Q/K), discards default leaf scores, and replaces them with a KRR or MLP-learned value representation (V). The final prediction is a dot-product Q·K·V aggregation mimicking a zero-temperature attention head, front-load

Design a neuro-symbolic memory architecture for autonomous agricultural agents, inspired by the brain's Complementary Learning Systems (CLS). The architecture decouples memory into a fast-learning episodic Knowledge Graph for temporal/spatial sequences (crop cycles, disease spread) and a slow-learning semantic ML layer that compresses mature data i

Design a deep multimodal emotion recognition (MER) architecture fusing speech (acoustic/prosodic features), facial expression video frames, and transcribed text for real-time mental health monitoring. The architecture will use modality-specific encoders (CNN/transformer for audio and video, pretrained LM for text) with a cross-modal attention fusio