Making own deep learning framework Fiera (https://github.com/xyzunreal/Fiera) for edge devices.* Quantization and Pruning of Network to reduce Network latency in the order of magnitude without significantly hurting accuracy. * Worked on different formats like HDF5, JSON, CSV, Protobuff, Flatbuffers for serializing and deserializing weights and network info. * im2col, Winograd and FFT algorithms to further speed up the convolution layer which is usually the bottleneck in Neural… Show more Making own deep learning framework Fiera (https://github.com/xyzunreal/Fiera) for edge devices.* Quantization and Pruning of Network to reduce Network latency in the order of magnitude without significantly hurting accuracy. * Worked on different formats like HDF5, JSON, CSV, Protobuff, Flatbuffers for serializing and deserializing weights and network info. * im2col, Winograd and FFT algorithms to further speed up the convolution layer which is usually the bottleneck in Neural Nets.* Proper memory layout of multidimensional matrices to maximize CPU cache usage, vectorization and make good use of scientific computing libraries like OpenBLAS. * Binarizing large FP32 matrices and packed them into int64_t. This reduced storage of weights by a factor of 32. Now, weights can be stored on SRAM instead of DRAM which is much faster and 100x more power-efficient. Show less

ResponsibilitiesChatbot for MereExams site to provide info about 40k+ colleges and 25k+ courses.College Predictor for given rank.Cloud server handling.Data cleaning, Analysis and Plotting.AchievementsIncreased web scrapping speed by 3-4 folds using threading.Reduced complexity from O(n^2) to O(n) of data mapping.Improved Context handling for chatbot by saving user session data on the server.