Machine Leaning Engineer Intern  @ Vosyn AI 

Role Overview:

• Fine-tuning the DeepSeek 8B language model using LoRA, QLoRA, and 8-bit quantization for Arabic–English and English–Japanese translation

• Utilizing PyTorch and Hugging Face Transformers for efficient multi-GPU training and checkpoint management.

• Working on building event-driven deployment pipelines on Google Cloud:

  • Model hosted on Vertex AI Endpoints.

  • Inference via Google Kubernetes Engine (GKE) and Cloud Run.

  • Triggers linked to Cloud Storage bucket uploads.

• Parallelizing inference using ThreadPoolExecutor, batch chunking, and caching strategies to support scalable translation of large files (e.g., SRTs).

•  Currently working on the deployment of a low-latency speech-to-speech (S2S) translation system, integrating a Node.js backend with a Django microservice, containerized using Docker and deployed on Cloud Run for scalable serving.

• Ugarding towards Compute Engines to maintain a monolithic architecture for real-time synchronization and optimized coordination across backend components with REST APis/ gRPCs.
  
  

Impact Highlights:

• BLEU improvement: 0.00 → 16.06 (JA→EN), 3.08 → 34.56 (AR→EN).

• BLEURT boost: +36%.

• Pipeline latency reduced from 2 minutes → 30–40 seconds for 25-block SRT files.

Tech Stack:

DeepSeek 8B, LoRA, QLoRA, 8-bit quantization, PyTorch, Hugging Face, Vertex AI, GKE, Cloud Run, Cloud Storage, Docker, GitHub Actions, Python,  Node JS, Django, CUDA, GCP.

 Growth & Takeaways:

• Strengthened expertise in efficient LLM tuning and deployment orchestration using event-driven cloud infrastructure.

• Improved understanding of event-based architectures, LLM I/O bottlenecks, and efficient parallel compute management

• Gained deeper insight into MLOps orchestration, parallel processing, and real-time system optimization and low-latency inference optimization.

Instructor Assistant Upward Bound  @ CU boulder

Role Overview:

• Supporting a programming fundamentals course with 15+ students.

• Assisting lab sessions on Python, covering core topics like:

  • Data types, control flow, and function design

  • File I/O operations

  • Debugging strategies and exception handling

• Providing personalised assistance via Jupyter Notebooks and live feedback during labs.

• Tracking student progress through performance metrics, surveys, and assignment reviews.

Impact Highlights:

• Increased student comprehension by 40% via personalised guidance and iterative feedback.

• Helped multiple students move from basic to intermediate proficiency in Python within one semester.

Tech Stack:

Python, Terminal, VSCode,  Excel,  Google Classroom.

Growth & Takeaways:

• Improved ability to explain abstract programming concepts through analogies and live examples.

• Learned how to tailor technical communication to varied learning styles and skill levels.

Machine Leaning Research Intern  @ DRDO INMAS

Role Overview:

• Designed and implemented a real-time video acquisition pipeline using:

  • Arduino UNO for hardware trigger generation

  • OpenCV and Python for real-time video capture and frame processing

• Achieved hardware-software synchronisation through serial communication and multithreaded signal handling.

• Applied advanced computer vision techniques:

  • Gaussian blur, adaptive thresholding, contour detection

  • Region of Interest (ROI) segmentation for feature extraction

• Enhanced the visual analytics pipeline for researchers performing object-based study of frames in lab settings.

Impact Highlights:

• Latency improvement: 73% faster retrieval.

• Robustified experimental imaging pipelines for extended-duration trials

Tech Stack:

Arduino, OpenCV, Python, Serial Protocols, Multithreading, USB Camera Interface, NumPy, Matplotlib, CNN.

Growth & Takeaways:

• Developed real-time embedded-vision system skills

• Learned performance tuning and multithreaded control at the edge device level

Machine Leaning Research Intern  @ DRDO INMAS 

Role Overview:

• Built CNN models using TensorFlow and Keras to detect COVID-19 from chest X-ray images.

• Preprocessed 700+ real medical images using:

  • noise filtering

  • Data augmentation (flips, rotations, scaling)

• Designed both binary and multi-class pipelines to analyze diagnostic trade-offs.

• Tuned hyperparameters using early stopping, dropout, and L2 regularization to avoid overfitting.

• Generated heatmaps for model interpretability, aiding physician validation of predictions.

Impact Highlights:

• Binary model achieved 93.75% accuracy vs. 71% for multi-class

• Produced visual activation maps for clinical feedback and explainability

Tech Stack:

TensorFlow, Keras, Python, Scikit-learn, Matplotlib, Grad-CAM, NumPy, OpenCV, Jupyter

Growth & Takeaways:

• Strengthened understanding of medical AI workflows, data bias handling.

• Learned how to align ML pipelines with real-world diagnostic requirements.