AYUSH KUMAR SAMAL

I’m an AI-focused B.Tech student building deep learning and ML systems that are both high-performing and practical to deploy. Through my coursework and hands-on work, I gravitate toward architectures, optimization, and end-to-end pipelines.

As a Deep Learning Research Intern at NIT Rourkela, I designed and developed a custom deep learning architecture for static hand gesture recognition. I incorporated residual connections, multi-scale feature extraction, self-attention mechanisms, and attention-layer regularization, then improved it beyond baselines. I also built end-to-end data pipelines and evaluated deployment on NVIDIA Jetson Nano with a focus on latency and resource efficiency.

My projects extend this systems mindset: I built an end-to-end ML compiler that translates high-level model definitions into optimized C code via a structured compilation pipeline. I also implemented and extended a Triton-based FlashAttention (forward/backward kernels, PyTorch autograd integration, and kernel autotuning), plus benchmarking and correctness validation against PyTorch SDPA. For applied safety work, I designed an autonomous agentic honeypot for scam detection and extraction using LangGraph workflows and FastAPI services.

I’ve strengthened my foundation with an Applied Machine Learning course and consistently validate ideas through experiments, benchmarking, and iteration. I’ve also earned recognition at the IndiaAI Impact Hackathon (Top 25 out of 800+ teams) and contributed as an organizer for HackNITR.