Latest news and updates from DSML Kazakhstan community
Stay up to date with the latest events
Congratulations to our community member Issagali Konysbayev, who developed a unique resume-based job recommendation system using advanced NLP models.
Working with IT specialists and analysts from the Workforce Development Center, the system was adapted and deployed on enbek.kz.
The service is built with FastAPI, sentence_transformers, and PyTorch. Its core is an NLP model that converts resume and vacancy text into embedding vectors. After jobs are filtered by parameters such as region or profession, these vectors are used to calculate cosine distance between vacancy vectors and the resume query vector. The system selects the nearest vectors by cosine similarity.
A distinctive feature is that the system learns not only from current occupations, but also from the semantics of an applicant's education, skills, and previous work experience. Issagali used a checkpoint from the middle of training rather than the final model to avoid overfitting solely to current occupations.
Recommendations are fast because the system uses only a retriever model, without an additional reranker classification model. The model was trained with ContrastiveLoss. Although Issagali did not have time to test other loss functions, there remains substantial potential for further improvement.
According to a Workforce Development Center press release, the project both demonstrates important innovation in recruitment and opens new possibilities for applying NLP to real labor-market problems.
The tool manages clusters while supporting scalability and fault tolerance.
The library helps you use a GPU cluster efficiently to build systems comparable in power to ChatGPT.
Read the blog post here.
Wishing everyone successful experiments! 🚀
CerebraAI, which develops an AI-powered solution for early stroke detection, has successfully closed its Seed 2 funding round.
The round was led by venture fund Big Sky Capital. Other investors included Most Ventures, Aloqa Ventures, Uzcard Ventures, and private investors including Timur Turlov, UMAY Angels, and Ynta Group. Most investors agreed that the startup's final valuation would be determined within three years, while others accepted the current pre-money valuation of $12 million.
Adil Nurgozhin, Managing Partner at Big Sky Capital, expressed confidence in CerebraAI's global potential and a willingness to support the startup on its path to success in the U.S. market.
On behalf of the community, we congratulate everyone involved, including the data science team led by Zhuldyz-Zhan Sagimbayev.
The annual Alem Cup 2023 programming championship began on September 8. It is organized by alem school with support from the Astana Hub international technology park. Participants compete for cash prizes by creating bots that battle other players in a computer game. People of any age and background can take part.
The prize pool is:
Your code will run every 15 minutes against code from participants at a similar skill level. A win adds ranking points and a loss removes them. A new ranking is formed after each round of arena games. The three highest-ranked players advance to the final.
Supported programming languages: Python, C++, JS, Go, Pascal, Java, PHP, C, C#, Ruby, Kotlin, Blockly.
The championship runs until October 12.
The 24th International Conference on Digital Signal Processing (DSP2023) took place in Greece this summer, featuring papers on digital signal processing and machine learning. DSML KZ members Zangar Yermaganbet @goofymonarch and @Ayana_Mussabayeva, together with NU professors, presented a paper on a double-input CNN used in spelling systems for people with amyotrophic lateral sclerosis.
A P300 Speller flashes symbols on a screen. When the user sees the symbol they want to type light up, it is treated as the target response. The task becomes binary classification: either the signal is a target and the user wants to type the highlighted symbol, or it is not.
The authors generate two types of spectrograms from EEG brain signals using independent component analysis and the Fourier transform, then pass the spectrograms through a double-input CNN.
The reported results show that the method is more effective than classical classifiers used in BCI systems, such as SVM and LDA.
The full paper is openly available on ResearchGate.
