Dmitry Kalashnikov

Experience

• 

  Penza Scientific Research Electrotechnical Institute

  Feb 2013 - Sept 2022

  • Implemented and improved the biometric voice authentication system, enhancing accuracy by 40% and decreasing error rates by 35%, utilizing TensorFlow, Keras, and OpenCV; • Developed a demonstration software based on human identification by signature. The final probability of an error of the first type is 0.01, of the second type is 0.08. In this work, I modernized the neural network of the Russian standard GOST R 52633;• Adapted voice model of speech translation on a laryngophone. This made it possible to use voice control for more than 250 people in the field;• Built a test biometric lock using 2D technologies of face, voice, finger; The introduction of technology into the university checkpoint reduced the number of unauthorized visitors by 40%;• Implemented my own voice authentication system into the general SDK of the Russian standard for biometric recognition technologies (faces, fingers, palm veins, voices, signatures). The strength of the biometric key has increased by 18-23 bits. The probability of a stranger passing through is 0.001%. The total probability of hacking the biometric system was 10^(-13). Show less • Created noise filtering algorithms for the voice biometric authentication system. The algorithm allowed the system to operate in high noise conditions with high recognition quality. Signal to noise ratio = 1/1, which is 65% higher than previous system capabilities.• Developed biometric tokens for the identification of medical workers. The results reduced the operating time of the access control system by 15% and reduced the incidence of unauthorized entry by 6%.• Processed raw face database of 150000 images data to train a neural network; • Researched and created software for finding malicious hacking signals obtained from an oscilloscope, using deep neural network.• Developed algorithms for compression and speech synthesis to create a new class of biometric vocoders, which speed was 300b/sec, which is 4 times faster than existing technologies.• Developed algorithms for recognition of printed Russian text, using Natural Language Processing with a particular emphasis on Large Language Models (LLMs) and Conversational AI. The final recognition quality on a test sample of 15,000 pages of text was 98.7%. Show less

  • Senior scientific researcher/Machine Learning Engineer

    Sept 2017 - Sept 2022

  • Software Engineer

    Feb 2013 - Sept 2017
• 

  UBPS

  May 2018 - May 2019

  Software Developer

  • Managed the integration of multiple biometric technologies into a unified SDK, coordinating with a core team of 16 and external stakeholders, achieving a 25% improvement in system efficiency.

• 

  Smart technology laboratory

  Oct 2022 - now

  AI/ML Development

  • Increased the accuracy of localization of identified defects in medical products by 20%. The final error as a result of the work performed was 0.8%, using PyTorch models;• Optimized the Differential Evolution of finding the global minimum of Rastring test function algorithm for a medical corporation. Implemented an intermediate algorithm from matlab to C++, increasing the processing speed by approximately 9 times. Reduced processing time from 2.5 hours to 15 minutes;• Modified the neural network algorithm for unmanned patrol drones, providing the ability to detect objects ranging in size from 10 to 40 pixels without loss of recognition quality, which is 70% less than initially possible. Show less