Ivan Karlov
#
karlovivan679@gmail.com |
+7 (996) 445-05-63 |
}
@IvanK2003
Summary
Data Scientist with 3.5 years of experience in ML and analytics. I design, ship to production, and
maintain data-driven features that automate processes and improve business metrics.
Experience
Data Scientist, IndorSoft February 2022 – June 2025
• IndorIntensity — automated traffic counting and reporting (project);
– Pipeline: YOLO one-vs-rest (1 pod = 1 class) → tracking (ByteTrack) → aggregation; 13 vehicle
classes + pedestrians; end-to-end export to report.
– Cost: consolidated inference pools and decoding (ffmpeg/NVDEC) → GPU-hours –35–45%.
– Scene calibration: lanes/crosswalk + homography; georeferencing ≤ 0.5 m (median, per ROI).
– Routing / OD: rule-based router on an intersection graph → route accuracy ≥ 0.95; flow-level
accuracy +25–35%.
– Conflicts (near-miss): TTC/PET from trajectories → precision 0.82–0.88, recall 0.75–0.83;
pilot identified 3 accident-prone hotspots.
– Optimization: INT8 + pruning + TensorRT/ONNX, async decode→infer→post → p95 <45
ms/frame (1 stream), 90–120 FPS on RTX 3060 (4–6 streams), CPU 1080p 18–25 FPS.
– Labeling: active learning (uncertainty + diversity) + pseudo-labeling → labeling costs –50–65%.
• Add-on to IndorTrafficPlan — detection and transfer of road signs from video to a road plan
(project).
– Detection: YOLO → F1 = 0.96, mAP@0.5:0.95 = 0.89.
– Classification: EfficientNet_V2 by sign groups → accuracy ≥ 0.98.
– Optimization & performance: hybrid MobileNet_V3 (tabular + visual features) → inference ×3
faster with no accuracy loss.
– Tracking: BoT-SORT + ReID → reduced ID-switches; robust to partial occlusions.
– Domains/synthetic: generation (night/rain/glare) + style transfer → dataset ×4; recall +8%,
IoU +12%, F1 +10%; QA via CVAT/FiftyOne.
– Geometry: distance triangulation to signs → error ±0.15 m (median, calibrated camera).
• Support ticket routing — CatBoost ensemble to redirect tickets to FAQ sections; team workload
–15%.
Data Scientist Intern, IndorSoft January 2022 – February 2022
• Wrote efficient SQL for data extraction and analysis; refactored legacy code.
• Sourced, analyzed, and prepared datasets; demoed results with Gradio.
• Built image clustering on ViT + K-means (Silhouette = 0.91 after PCA); improved YOLO classifier
by +12% F1 via a new split.
Skills
Languages/ML: Python, Pandas, NumPy, Scikit-learn, CatBoost, PyTorch, TensorFlow, ONNX
MLOps/Backend: Docker, Kubernetes (KEDA, Argo Rollouts), Airflow, FastAPI, Prometheus, Grafana
Data/DB/GIS: SQL, PostgreSQL, PostGIS
Other: Git, Linux, Plotly, Computer Vision, NLP, Data Mining, OOP
