Daniyar Bakir

A Python DRF and React-based CRM tool. The tool should help automate invoice generation and structure the internal processes of the business.

I am the sole developer working on this project. The back end is based on the Django REST Framework and PostgreSQL and utilizes the Service-Controller-Repository pattern. The front end is based on the React Refine framework. CI/CD and deployment are done via GitLab CI/CD and DigitalOcean Droplet.

An automation pipeline for processing the raw audio data for the Speech Data Annotation team. This tool speeds up the batch preparation process, eliminates human interaction, and visualizes the whole process.

I was the initiator and sole developer of this project, which sped up the annotation process from 3.5 – 4 hours per batch to 45 –50 minutes per batch. It was completed using Apache Airflow, FFmpeg, Docker, in-house STT (for pseudo annotation), Python, and Flask.

The project involved training time optimization of the speech-to-text (STT) pipeline by restructuring scripts and storage. The problems included increasing checking hypotheses on STT and reducing scheduling problems for the GPU server. As an ML engineer, I had to find optimization points and implement them, restructure the training pipeline, and reorganize the data storage structure. It resulted in a significant reduction in training time and the removal of idle GPU server operation during the initial CPU computation steps by segregating the training pipeline into two parts: CPU and GPU.

A virtual assistant can improve users' overall experience with everyday life tasks by completing simple tasks. The service needs the users' commands and completes their requests by passing them through a technical pipeline that consists of speech-to-text, natural language processing, and text-to-speech modules. I acted as an ML engineer, designing and implementing a specific speech-to-text system for the virtual assistant and implementing integration with the virtual assistant pipeline.

A web project for internal usage only. It was developed to replace a 3rd-party vendor project with an in-house customized application.

In this project, I was the sole developer responsible for the data scraping, back end, front end, CI/CD, and deployment.

A simple-to-use speech-to-text training pipeline and service for a 2-language environment, Kazakh and Russian. For example, the STT service can be retrained to match a particular vocabulary used in a financial call center. As an ML engineer, I researched and implemented various STT architectures, trained and validated models on new data, developed and improved the existing STT models, and optimized the data preparation pipeline.

The project's goal was to increase the client's involvement in the market, optimize the sales managers' working time, and increase the company's product presence among the businesses. I was an ML engineer contributing to designing a call center's pipeline that consisted of STT, NLP, TSS, telephony, and message queue. We also created an STT system suitable for the company's jargon. It resulted in a more than four-fold increase in cold calls.

It is an internal web tool for large-scale audio transcription. The audio transcription required a lot of human resources, and the existing open-source audio annotation tools were slow and difficult to use while requiring manipulation with the user's machine, that is, installation or sharing physical data. This project aimed to develop a web tool that would remove the large downsides of open-source tools by eliminating, distributing, and collecting data to and from annotators, introducing automatic data engineering, and customizing the tool for all annotators simultaneously.

As a full-stack developer, I designed and developed the architecture of the back-end system, data collection pipeline between the database and ML engineers, and UI. I also developed the front-end system and transferred the project to a new team.

The project involved developing an ML service to determine spoken language in an audio file by developing a spoken language identification model and integrating this model into the data preparation pipeline. As an ML engineer, I designed, implemented, validated, automated, and improved the language identification model in audio files. I significantly reduced the time spent annotating audio file language and prepared and curated audio datasets.

A fun service for users to find a celebrity face. For example, a user uploads their face and selects a target celebrity. The app will display the face of a celebrity that looks like the user, then the next face that looks like the previous celebrity until the target celebrity is shown. I acted as a back-end developer and designed a scalable microservice architecture connecting several ML microservices. I also developed the app's back end and front end and implemented a production-ready code.

The business idea of the project was to collect user data from external sources to improve the company's products, such as the recommendation system. As the back-end developer, I explored various ways to retrieve data bypassing pagination limitations, developed the API to update existing data in the database, and implemented a single fetch of many users' data.

A web project based on Python and Flask for a university ranking system. The business idea was to create an academic ranking system for universities and present it as a website. I acted as a full-stack developer managing data crawling for 4,000+ universities, developing the back-end and mobile-first front-end systems, creating an interactive map to visualize top universities per country, and deploying the project on DigitalOcean.