Skip to content
MM

Maksat Manapkhan

@maks

MLE in CV

bs in math at NU, mle at kaspi.kz in cv team, want to learn more interesting things, be a better person @monokorenko

Career

Experience: 1 years

Experience

Machine Learning Engineer

Kaspi.kz

Kazakhstan-based technology and financial services company.

Aug 2025 — Present

Built and deployed ML systems for moderation, OCR, and video matching; improved automation by 8.12% and delivered video-matching recall of up to 99.67%.

• VLM Verification Layer: Introduced vision-language models as an automated post-check over four production classifiers, filtering borderline predictions to cut false positives and raise the overall automation rate by +8.12%. • Chinese OCR Service (PP-OCRv5): Built and shipped the full production pipeline — preprocessing, text detection, angle classification, recognition and structured post-processing — served as a batched FastAPI microservice in Docker with logging and failure handling.

Research Assistant

Nazarbayev University

Research university in Kazakhstan.

Jul 2025 — Jan 2026

Implemented and validated an ADMM algorithm for constrained sparse group LASSO formulations in high-dimensional forecasting.

• Implemented the ADMM algorithm in R to optimize Multivariate Constrained Sparse Group LASSO formulations for high-dimensional forecasting, validating stability against analogous frameworks from the literature.

Computer Vision Engineer Intern

Scientia

Nov 2024 — Jan 2025

Developed computer vision methods for box-corner and fracture-edge detection, including box-corner detection at 0.06 MSE.

• Automated box corner detection to 0.06 MSE with a ResNet-34 trained on a quadrant-based image division strategy and custom augmentation pipelines. • Built an ellipse-fitting mechanism for fracture edge detection using OLS regression, improving geometric precision in structural analysis.

Research Assistant

Nazarbayev University

Research university in Kazakhstan.

Aug 2024 — May 2025

Developed physics-informed and convolutional neural networks to estimate acoustic source parameters, reaching 0.08 MSE.

• Developed a Physics-Informed Neural Network reconstructing acoustic source parameters, embedding the Helmholtz equation in the loss function to reach 0.08 MSE. • Designed 1D CNNs over multi-channel complex inputs (amplitude/phase) from microphone arrays, mapping sensor data to source parameters.

Education

Bachelor of Science, Mathematics

Nazarbayev University

2021 — 2025

Selected projects

Car Accident Detection

Extracted frame-by-frame vehicle bounding boxes with YOLOv11l to build time-series data, then trained an LSTM with logarithmic probability targets to localize impact timing.

Achievements

IEEE Hackathon — 3rd Place: Sergek Group case study

Skills

PythonRMatlabPyTorchPaddleOCRVLMsOpenCVScikit-learnFastAPIDockerGitLinux