Sukhmandeep Singh

• Built an automated end-to-end AI pipeline spanning data ingestion, model inference, candidate evaluation, and ranking, with validation, error handling, structured logging, and reproducible execution. • Converted CPU-bound protein structure-prediction workflows to GPU-accelerated pipelines using batched inference and multi-worker CPU parallelism, reducing processing latency and compute resource usage. • Deployed and orchestrated pipelines across AWS, SLURM, and Nextflow, containerized with Docker and version-controlled with Git, enabling reproducible runs across local, HPC, and cloud environments. • Developed and integrated protein language model and structure-prediction pipelines for AI-driven biomolecule discovery, applying transformer-based sequence modeling in an applied research workflow. • Built and maintained automated evaluation, filtering, and ranking pipelines that combine structural predictions with sequence- and structure-based scoring for generated candidates, in collaboration with computational biology and R&D teams.

• Deployed lightweight ML/DL models for fully offline classification and regression across Android, Raspberry Pi, and NVIDIA Jetson Nano devices; reduced model size 10MB → 3MB via quantization, achieving 230–350ms end-to-end inference latency. • Integrated models into a Flutter application with on-device SQLite caching, maintaining under 20MB memory and under 20% CPU utilization on Android. • Built Dockerized inference pipelines and mentored 10+ interns across ML, deep learning, IoT, and real-time sensor integration.

• Built end-to-end ML pipelines for fraud detection using XGBoost and Random Forest, achieving 92.4% accuracy (ROC-AUC: 0.76) through feature engineering, hyperparameter tuning, and cross-validation. • Scraped and cleaned large-scale datasets using BeautifulSoup, Selenium, and Pandas to build structured training sets.