Sahil Khan

Modernized data monitoring infrastructure and engineered advanced AI/ML systems for predictive fault detection and anomaly detection, significantly enhancing system observability and operational reliability.

• Modernized data monitoring by migrating from a cumbersome custom FastAPI dashboard to a scalable Grafana and InfluxDB stack, decoupling analytics from the production database, dramatically improving system observability.
• Engineered a predictive fault detection system using physics-based synthetic data generation, robust fault labeling, and custom LSTM model training, enabling reliable and scalable fault forecasting.
• Built an unsupervised operational anomaly detection pipeline with LSTM autoencoder, latent features, and clustering for proactive issue identification.
• Configured an alarm system using webhooks to deliver immediate updates on system alerts and operational faults, enhancing response times and system integrity.

Developed high-fidelity digital twin solutions for solar inverters, achieving over 95% real-time performance accuracy, and pioneered a RUL estimation model to advance predictive maintenance.

• Developed a high-fidelity digital twin for solar inverters, delivering over 95% real-time performance accuracy.
• Pioneered a RUL estimation model achieving 70% accuracy without failure data, significantly advancing predictive maintenance capabilities.
• Built FastAPI backend services to seamlessly stream digital twin and RUL estimation analytics on dashboards, enhancing data accessibility.

Conducted research and fine-tuned a Detection Transformer with ResNet50 on the WiderFace dataset, significantly improving accuracy on tiny face detection.

• Selected and fine-tuned a Detection Transformer with ResNet50 on the WiderFace dataset (Hard set), improving accuracy on tiny face detection.
• Improved DETR's detection capabilities via object query augmentation, tuning, data augmentation, and backbone optimization.

Recreated the TinyStories research by training and evaluating a compact Transformer on a dataset of children's stories.

Developed a Streamlit-based platform where an AI Tutor answers student queries, auto-generates quizzes, and provides tailored feedback.

Built a custom Agentic RAG with the Gemini model from scratch, incorporating Planning, Reasoning, Action, and Reflection.

Developed a probabilistic spelling corrector and an LSTM-based Neural network for next-word prediction.

Predicted the Remaining Useful Life (RUL) of IGBT modules in wind turbines as a B.Tech project.