Srijan Dokania

About

Built and deployed an Azure-based, fine-tuned CLIP visual retrieval system at Staples for a ~300k-item product catalog, improving edge-case recall by 12% by engineering a custom delta-similarity/dynamic-margin loss. Also has robotics experience using ROS2 for sensor/compute orchestration, including GPS-time-synchronized sensor triggering for robot swarms and latency-bounded optical-flow benchmarking for edge deployment.

Experience

Education

Key Strengths

• Shipped production visual search/retrieval system for ~300k-item catalog
• Improved edge-case recall by 12% via custom delta-similarity/dynamic-margin contrastive loss
• Strong understanding of hard negatives and training stability techniques (warmup + gradual loss weighting + low LR)
• Orchestrated end-to-end ML pipelines from ingestion to deployment using cloud tooling (Azure/AWS/PySpark)
• Resolved PySpark reliability issues from data skew using aggressive checkpointing
• Robotics orchestration expertise with ROS2 across sensors/compute/odometry
• Built GPS-time-synchronized distributed sensor triggering to mitigate network-latency sync issues in robot swarms
• Reliability-focused AI workflow design using modular/hybrid retrieval (keyword + semantic) to reduce errors
• Performance/latency-driven model evaluation (e.g., optical flow <90ms; benchmarking on KITTI) before edge deployment
• Effective collaboration with non-technical stakeholders (merchandisers) by translating qualitative feedback into measurable model improvements

Reference Highlights

• mature communicator
• explains complex concepts in simple, layman terms
• strong stakeholder management and communication
• able to translate data insights into clear visual explanations
• good at designing and shipping AI/LLM products
• effective at handling technical questions during workshops/demos