Design a CNN architecture augmented with L1 (Lasso) regularization that applies sparsity-inducing penalties on convolutional and fully-connected weights, producing a sparse feature hierarchy. The design covers integration via weight decay, structured sparsity variants (group lasso on filters), and a forward-pass formulation where lasso-proximal ste

Design a diffusion-based LLM architecture that replaces autoregressive next-token prediction with iterative denoising over discrete token sequences. The architecture uses a bidirectional transformer as the denoising backbone, a continuous or discrete diffusion process over token embeddings, classifier-free guidance for conditioning, and accelerated

Design a unified Latent Diffusion Model (LDM) for any-to-any imputation of missing glioma MRI sequences (T1, T1c, T2, FLAIR) across pre- and post-operative phases. The model will use combinatorial masking during training, pathology-aware loss functions to preserve tumor margins and textures, and produce stochastic uncertainty maps to quantify recon

Design a multi-shopper, multi-product autonomous checkout vision system that generalises a gesture-based prototype. The system integrates hand-item association (shopper-ID to item-ID), temporal cart-state modelling via a state machine or transformer, and robustness to occlusions, put-backs, and hand-offs. Evaluation targets on-device latency and pr

Design and build a privacy-preserving hate speech detection (HSD) system that operates agnostically to author identity. The model uses a Transformer backbone (e.g., BERT/RoBERTa) with differentially private training (DP-SGD/DP-Adam) and adversarial identity disentanglement to decouple hate-signal from author-specific stylistic fingerprints, maximiz

Design a family of SOTA CNN architectures (ConvNeXt/ResNeXt-style) for ESC-50 audio spectrogram classification, parameterized across 5M, 10M, 20M, and 30M sizes. The design uses mel-spectrogram inputs, depthwise separable convolutions, squeeze-and-excitation blocks, and proper regularization (SpecAugment, Mixup, Label Smoothing) to achieve strong g

Design an enhanced ConvNeXt V2 architecture incorporating multi-scale dynamic convolutions, adaptive feature recalibration via a lightweight coordinate-aware gating mechanism, and a refined pretraining strategy (improved FCMAE with spatial-aware masking and hierarchical reconstruction loss) to improve representation quality and downstream performan

Design an ML-enhanced receiver chain for IoT-to-LEO-satellite communication, where conventional signal-processing algorithms (e.g., channel estimation, synchronization, detection) are augmented or replaced by learned models under low-SNR, high-Doppler, and resource-constrained IoT conditions.

Design a generative AI pipeline for restoring missing sections of historical paintings while preserving artistic style. The system will combine StyleGAN2 for style-preserving inpainting with diffusion models for fine-grained detail recovery, and will incorporate explainability mechanisms (attention maps, CycleGAN before/after comparisons) and trust

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