Research & Publications
Peer-reviewed manuscripts, algorithmic architectures, statistical benchmarks, and open-access datasets in low-resource multilingual machine learning and adaptive sorting.
First Author, Advised by Dr. Osman Khalid — Manuscript Submitted
- Engineered a novel deterministic, comparison-based sorting algorithm in C++ that dynamically adapts to $K$, the minimum number of nondecreasing subsequences, formally proving algorithmic optimality where $K = \mathrm{LDS}(A)$.
- Designed a robust two-phase architecture: a stateful, hinted greedy decomposition utilizing exponential galloping, followed by a dynamically selected $K$-way loser-tree tournament or two-way galloping merge strategy.
- Benchmarked against
std::sort and TimSort on Apple M1 architectures handling datasets up to 100M elements; achieved up to 1.45× wall-clock speedups on real-world GitHub Archive event streams ($N=55\text{K}, K=368$) and consistent 30–50% performance gains on synthetic interleaved stream workloads.
C++
Algorithms
Optimal Sorting
Tournament Tree
Galloping Search
Benchmarking
First Author
- Architected a comprehensive multi-dialect evaluation framework ($n=5,207$ tweets via ArSarcasm corpus) benchmarking AraBERT, MARBERT, and CAMeL-BERT across 5 distinct regional varieties (Egyptian, Gulf, Levantine, Maghrebi, MSA).
- Formulated the variance-based Dialect Robustness Index (DRI) to quantify structural resilience; validated empirical findings using 1,000-resample bootstrap confidence intervals and instance-level paired McNemar tests ($p<0.001$).
- Identified and resolved a critical pipeline memory artifact in zero-shot checkpoint configurations, proving empirically that fine-tuned task adaptation significantly supersedes frozen-representation separability in PLMs.
PyTorch
Hugging Face
AraBERT
MARBERT
Dialect Robustness Index (DRI)
Statistical Significance
First Author
- Benchmarked three multilingual transformer encoders (XLM-R, mBERT, MuRIL) across 54 controlled training runs on 28-class Roman Urdu emotion classification, sweeping preprocessing regime (OFF/ON), spelling-variant augmentation ($k \in \{0,1,3\}$), and epochs (2–4) under frozen, leakage-free splits.
- Achieved the best fine-grained result of Macro-F1 = 0.2869 (XLM-R, ON, $k=3$); ran a size-controlled ablation isolating the contribution of orthographic variant exposure from raw training-data volume.
- Standardized zero-shot baselines by mapping the GoEmotions taxonomy to Ekman-6, attaining 70.31% accuracy (Macro-F1 = 0.5004) with XLM-R, while generating publication-grade analytical artifacts (confusion matrices, per-class reports, run logs).
Transformers
XLM-RoBERTa
mBERT
MuRIL
GoEmotions
Roman Urdu
Ablation Studies
- Engineered a highly fault-tolerant English → Roman Urdu translation pipeline leveraging Ollama Gemma-3 (4.3B parameter LLM) to construct the largest public Roman Urdu emotion dataset (54,263 samples) tailored for low-resource NLP research.
- Developed a resumable, chunked-processing backend (100-row batch execution) enforcing strict emotion-preserving constraints, automated quality control, and anomalous audit flagging to safely operate within Kaggle/Colab transient limits.
Zenodo
Ollama Gemma-3
Dataset Creation
Low-Resource NLP
Data Engineering