Margarita Latushkina

Applied AI Scientist / Research Engineer

Zoolatech

Apr 2025 - Jan 2026 (10 months)

• Fashion-LLM Project (Multimodal LLMs & VLMS)
Ο Large-scale fine-tuning of multimodal LLMS for fashion recommendation (text + image embeddings).
Ο Designed and implemented custom embedding heads (CLIP-style) for efficient retrieval.
Ο Built end-to-end multimodal training pipelines, integrating text, image, and metadata, optimized for GPU throughput and convergence.
• Experimented with hyperparameters, data mixtures, and model architectures to improve retrieval and recommendation quality.
Ο Practical experience with single-GPU fine-tuning for prototypes and iterative experiments.
Ο Distributed training with DeepSpeed.
Key Skills: multimodal model training, LLM/VLM fine-tuning, distributed training, pipelin

Data Scientist

SayGames

Sep 2024 - Feb 2025 (6 months)

• Predictive analytics
Ο Developed predictive models for ad performance using video embeddings (ViViT) and structured/tabular features (TABM/XGBoost).
Ο Built text embeddings for classification and sentiment analysis pipelines.
• Optimized data preprocessing and model inference for experimental pipelines.
Key Skills: ML pipeline design, embeddings, video analytics, inference optimization.

Machine Learning Engineer

Quantum

Aug 2022 - Jun 2024 (1 year 11 months)

• Satellite Image & Object Detection Pipelines
Ο Built semantic segmentation models (U-Net variants) for cloud, shadow, and carbon detection.
Ο Developed object detection pipelines using YOLO models for satellite imagery.
Ο Experience with distributed training frameworks for CV tasks, including DDP/FSDP concepts.
• Optimized inference throughput and deployed pipelines for large-scale satellite data.
• NLP & LLM Applications
Ο Fine-tuned BERT-based models for sentiment analysis, summarization, and customer feedback.
Ο Built RAG-based chatbots integrating LLMs with retrieval pipelines.
Ο Research-focused exploration of text embeddings, knowledge-aware retrieval, and optimization strategies.
Key Ski

Deep Learning Engineer

Oxagile

Aug 2019 - Aug 2022 (3 years 1 month)

• Computer Vision
Ο Implemented object detection, face recognition, tracking, and speed-optimized inference solutions. Used ResNet, ArcFace, DaSiamRPN.
• Access Control System
Ο Developed ML models for safety and security applications. Used ncnn, onnx, TensorRT for optimization on edge devices.
• Speech Analytics
Ο Built deep learning models for emotion recognition and keyword detection in speech. Applied DeepSpeech, Wav2Vec.

Margarita Latushkina

About

Detail

Timeline

Résumé

Margarita Latushkina

About

Detail

Behavioral traits