Research
Selected publications
Preprints, benchmark datasets, and applied work in AI systems, data infrastructure, and privacy-preserving computation.
A Unified Denoising and Adaptation Framework for Self-Supervised Bengali Dialectal ASR
First-author preprint with Imran and Tuhin Sheikh. Bengali has over 270 million speakers and wide dialectal variation, and field recordings are noisy. Our framework builds on WavLM, whose masked speech denoising pre-training already resists acoustic distortion, and adds a two-stage fine-tuning strategy: adapt to general Bengali first, then specialize for dialectal recognition under noise with targeted augmentation. Tested across multiple dialects from clean audio down to low signal-to-noise ratios, it beats fine-tuned wav2vec 2.0 and large multilingual Whisper, setting a new state of the art for the task.
An Empirical Benchmark Dataset for Paillier-Based Privacy-Preserving REST API Gateways
The homomorphic_request benchmark dataset for Paillier-based homomorphic encryption in privacy-preserving REST API gateways. It covers concurrency scaling, batch variation, homomorphic overhead, burst traffic, and fault injection, with per-request telemetry and a reproducible Go harness. Daffodil International University. DOI: 10.5281/zenodo.18655966
Experiments
Engineering case studies
Benchmarks, prototypes, and write-ups from my own experiments. The code and data are on GitHub.