.png&w=256&q=75)
Yashwin Sharma@sharmayashwin0116133
1
Events attended
1
Submissions made
India
1 year of experience
About me
AI and Data Science engineering student passionate about deep learning, edge computing, and hardware-software optimization. I specialize in building high-performance, privacy-preserving machine learning pipelines with Python, PyTorch, and OpenCV, with a strong focus on maximizing local hardware performance. Committed to bridging the gap between advanced AI architectures and practical, on-premise execution.
🤓 Latest Submissions
Wafer Defect Copilot: AI Yield Optimization
The Problem In semiconductor manufacturing, a single 3nm silicon wafer costs upwards of $19,500. When a physical defect occurs on the line, traditional inspection systems generate massive false positives. Engineers waste hours manually searching through (SOPs) to find the root cause. Meanwhile, machinery continues to process compromised silicon, which compounds not only the financial waste but also a massive environmental toll, silently consuming vast amounts of electricity, toxic chemicals, and thousands of gallons of pure water for ruined yields. The Solution Wafer Defect Copilot is a closed-loop, AI-powered assistant designed for night-shift fab engineers. It processes spatial wafer maps to instantly classify defect patterns (like Edge-Ring or Scratch) using a custom Convolutional Neural Network. If a critical threshold is met, it triggers a RAG pipeline utilizing sentence embeddings and a locally hosted Large Language Model to retrieve the correct equipment manual and generate a cited, step-by-step troubleshooting guide. AMD Edge-to-Cloud Integration To meet the extreme throughput demands of a factory floor, this entire pipeline was built and trained natively on AMD infrastructure. Using the AMD Developer Cloud, the CNN was trained end-to-end on an AMD Radeon Pro W7900D using the ROCm PyTorch stack (converging in just 184 seconds with an F1 score of 0.86). The inference engine runs the highly quantized Qwen2.5 7B LLM natively on the AMD GPU, generating structured engineering recommendations at 30.3 tokens per second. Environmental Impact By shifting defect resolution from an end-of-line manual chore to an automated AMD-powered pipeline, fabs can intercept excursions early. This prevents millions of dollars in scrapped silicon, drastically reduces machine downtime, and halts the hazardous material and energy waste associated with processing dead wafers, proving that intelligent automation is a requirement for sustainable manufacturing.
18 Oct 2026
