Implemented a Persian grapheme to phoneme model with a sequence to sequence transformer. Two variations of the model were trained: one with 3 encoders and 3 decoders and another with 5 encoders and 5 decoders. The larger model achieved a word error rate of 26.5% and a phonetic error rate of 6.19% on the test set. Source code and checkpoints are available at:
Project LinkUtilized IPOPT optimizer and pyomo library in order to find different 2x2 matrix multiplication algorithms with 7 scalar multiplications. This work was inspired by the AlphaTensor paper. Source code available at:
Project LinkImplemented a search engine using an inverted index and basic techniques. Enhanced the search engine by incorporating the tf-idf method, improving ranking of search results. Implemented index elimination and champions list techniques for optimization. Utilized Elasticsearch to expand capabilities, including spell correction, boolean query handling, and k-nearest neighbor algorithm for more accurate results. Source code available at:
Project Link