Karan Variyambat

About

Developed a production wildfire smoke detection system where smoke is visually subtle and easily confused with fog/clouds; addressed this with a hybrid CNN+LSTM+ViT model and multimodal weather features to reduce false positives. Experienced running scalable, reproducible ML pipelines on shared GPU infrastructure using Slurm and Kubernetes-style batch jobs with checkpointing, retries, and rigorous error analysis.

Experience

Education

Key Strengths

• Built and deployed a production wildfire smoke detection system from camera imagery
• Reduced false positives by combining visual modeling with contextual weather signals (multimodal approach)
• Designed hybrid deep learning architecture (CNN + LSTM + Vision Transformer) for subtle, ambiguous visual detection
• Handled real-world data issues: temporal misalignment and missing values across modalities
• Scaled and stabilized ML training/eval via Slurm and Kubernetes-style batch jobs with checkpointing and retries
• Strong evaluation discipline: metrics defined upfront, offline benchmarks, end-to-end stability checks, and scenario-based error slicing