Questions worth asking.

We introduce a 3-bit activation and weight quantization framework for low-resource Bengali transformer encoders. Our approach retains 97.4% F1-score on sentiment and named entity recognition while reducing memory footprint by 78%, enabling on-device inference on affordable Android micro-architectures.

A lightweight zk-SNARK cryptographic aggregation protocol tailored for distributed IoT telemetry sensors. Mitigates Byzantine gradient poisoning attacks while preserving complete non-invertible patient privacy without requiring heavy homomorphic encryption overhead.

We formulate multi-tier microservice cache eviction as a dynamic bipartite dependency graph and propose GNN-Cache, an asynchronous neural eviction predictor yielding 22% higher hit-rates over LRU/LFU baselines under bursty web traffic distributions.