Hallo world, I'm Muhtasham passionate MLE
Field of interests: LMOps(MLOPs for Language Models), Effeicient Transfer Learning, Out-of-Distribution Detection, Synthetic Data Augmentation, NLP and ASR for Low-Resource Languages
Skills 🛠️
- Languages: Python, C++, Matlab, Java
- DS/ML/DL: PyTorch, Tensorflow, Hugging Face
🤗 , Triton - RDBMS: NoSQL
- Big Data: Hadoop Spark
- DevOps: Linux, Git, Docker, REST API, Azure DevOps
Work experience 👔
| Job Position | Company | Field | Work Period |
|---|---|---|---|
| ML Engineer | Munich RE | ML, Model Operationalization, LLM | 15.11.2022 — Current |
| ML Engineer | Alyne | NLP, Multi-Task Learning | 01.03.2022 — 31.08.2022 |
| ML Engineer | Datamics | ML, Anomaly and Out of Distibution Detection | 15.01.2021 — 28.02.2022 |
| ML Engineer | Mirage | CV, Synthetic Data Augmentation | 01.02.2020 — 01.08.2020 |
| ML Engineer | Vispera | CV, GANS | 01.07.2019 — 15.09.2019 |
Education 🎓
- Master’s degree at Technical University of Munich (2020 - Sep 2022)
- Sequential Multi-task Learning in NLP with Focus on RegTech Domain
- Bachelor's degree at Sabanci University (2016 - 2020)
- Diploma thesis: "A Deep Neural Network for SSVEP-based
🧠 Computer Interfaces" - Featured in ML Monthly YouTube Show
- Diploma thesis: "A Deep Neural Network for SSVEP-based
Projects 🚀
-
📑 StarCoder- May the Source be with you - The StarCoder models are 15.5B parameter models trained on 80+ programming languages from The Stack (v1.2), with opt-out requests excluded. The model uses Multi Query Attention, a context window of 8192 tokens, and was trained using the Fill-in-the-Middle objective on 1 trillion tokens. -
📑 ECOC based Multi-Class Classification for BCI - Developed DNN, under ECOC framework to detect the character out of 40 possibilities that the (disabled) subject wants to communicate via Brain Computer Interface (BCI). -
📑 Fault-Tolerant Strassen-Like Matrix Multiplication - Designed parity calculations to make matrix multiplication calculations, using Strassen-Like algorithms on cloud computing environment (with parallel setting) fault-tolerant. -
📑 TajBERTo - Developed first ever RoBERTa-like language model based on Tajik language after extensive filtering of the dataset. The model can be used for masked text generation or fine-tune it to a downstream task.