Maria Petrovskaya
Data Pipeline & Feature Engineering: 5/10
Model Architecture & Training: 4/10
Experimentation & Evaluation: 4/10
Active 18 days ago
Invite to interview
Message
Download CVCV
Overview
Technical skills
Roles
Overview
A practical ML engineer at a middle level who independently builds end-to-end experiments in NLP and medical image/time-series classification. The strongest proven skill is applied model development and evaluation, evidenced by the SpiroCNN architecture, training/evaluation loop and dataset handling in the spirometry notebook. There is limited evidence of production-grade MLOps, automated experiment tracking, unit tests, or rigorous distributed/efficiency engineering.
Technical skills
AI/ML
Gensim
TF-Keras
CatBoost
Scikit-learn
Pandas
NumPy
spaCy
NLTK
scikit-image
Pillow
Torchvision
Analytics
Matplotlib
Seaborn
Middle AI/ML Engineer
Confidence: High ML Engineer
A practical ML engineer at a middle level who independently builds end-to-end experiments in NLP and medical image/time-series classification. The strongest proven skill is applied model development and evaluation, evidenced by the SpiroCNN architecture, training/evaluation loop and dataset handling in the spirometry notebook. There is limited evidence of production-grade MLOps, automated experiment tracking, unit tests, or rigorous distributed/efficiency engineering.
Model Architecture & Training
4/10
How well models are designed and trained
Custom model definitions and end-to-end training were implemented in both PyTorch and TensorFlow, including a bespoke 1D CNN (SpiroCNN) for time-series derived from images and a TF Keras embedding-based text classifier. Training loops, loss choices and optimizer usage are explicit, but there is limited evidence of advanced training engineering (distributed training, strong regularization ablations, or rigorous hyperparameter scheduling).
Evidence
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: class SpiroCNN and train_model() implementation
Classification of Text Complexity W2V and TensorFlow/Vers2_Classification_of_EnglishText_Complexity_Word2Vec_Tensorflow.ipynb: TensorFlow model with embedding layer and use of pretrained Word2Vec weights
Data Pipeline & Feature Engineering
5/10
How data is prepared for models
Concrete data-loading and preprocessing pipelines are present for multiple domains (subtitle parsing, PDF dictionary extraction, image-to-time-series conversion, HOG features). There is clear feature engineering (lemma-level dictionary percentages, TF/TF-IDF pipelines, GAF/HOG strategies), but production-grade data validation and leakage guards are limited.
Evidence
Natural-Language-Processing/Text complexity levels/Classification_of_text_complexity_levels.ipynb: load_subs(), clean_subs(), df_to_list() and dictionary-based feature creation
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: MedicalDataset class, _image_to_time_series(), HOG usage and image file path handling
Experimentation & Evaluation
4/10
How results are measured and tested
Reasonable experimentation and evaluation practices - cross-validation, GridSearchCV, stratified splits, CatBoost feature importance, and multiple metric reporting (F1, ROC-AUC, confusion matrices). Notebooks include plots and saved metadata, but there is no evidence of experiment tracking integration (W&B/MLflow) or reproducible run configs beyond notebook cells.
Evidence
Classification_of_text_complexity_levels.ipynb: GridSearchCV over pipelines with StratifiedShuffleSplit and f1_weighted scoring
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: evaluate_model(), confusion matrix, ROC-AUC reporting and model_metadata_7.json saving
MLOps & Deployment
2/10
How models are shipped to production
Basic model serialization and metadata saving (pickle, torch.save, zipping artifacts) and a safe-loading wrapper are present. There is no evidence of deployment, serving, CI/CD, model versioning, monitoring, or drift detection for production usage.
Evidence
Classification_of_text_complexity_levels.ipynb: pickle.dump(pkl_filename) saving of classifier pipeline
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: torch.save checkpoint, model_package_7.zip creation and load_model_safely()
Computational Efficiency
2/10
How efficiently computing resources are used
Some attention to efficiency like using trainable=False for pretrained embeddings and reasonable batching, but no profiling, GPU/CPU performance tuning, quantization, or distributed training evidence is present.
Evidence
Vers2_Classification_of_EnglishText_Complexity_Word2Vec_Tensorflow.ipynb: embedding layer using pretrained Word2Vec with trainable=False
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: use of DataLoader and batch training in train_model()
Research Depth & Innovation
2/10
Depth of research and new ideas
Shows applied creativity (image-to-time-series conversion, exploring spiral scans, HOG and Gramian-Angular-Field ideas) and reproduction of paper-style pipelines, but there is no clear evidence of novel algorithms, rigorous ablation studies, or peer-reviewed research standards.
Evidence
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: image_to_time_series, spiral_scan placeholders and multiple transformation strategies
Vers2_Classification_of_EnglishText_Complexity_Word2Vec_Tensorflow.ipynb: training Word2Vec embeddings and integrating them into a TF model
Expertise
Medical AI & Healthcare• Middle
Conversational AI & Chatbots• Middle
Industries
Education• Middle
Health Care• Middle
Recommendations
- Lead prototyping and PoC work for NLP classification or language-education features (text preprocessing, embeddings, pipeline tuning).
- Develop clinical ML prototypes for medical signal/image classification where data is modest and custom preprocessing (image-to-time-series, HOG) is required.
- Extend existing notebooks into reproducible training scripts with experiment tracking (W&B or MLflow), unit tests, and CI to move models toward production.
- Harden data pipelines and add validation checks, data schemas and privacy-aware handling for medical datasets before deployment.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Data Scientist
Confidence: High ML Practitioner
A pragmatic machine-learning practitioner at a solid middle level who builds end-to-end experimental projects and prototypes. The strongest proven skill is applied model development and evaluation for NLP and small medical imaging tasks - evidenced by the text classification pipeline (Count/Tfidf + GridSearchCV + CatBoost) and the SpiroCNN training/evaluation notebook. The public artifacts show limited productionization, sparse reproducibility tooling (no pinned env or CI), and little statistical formalism beyond standard metrics.
Statistical Rigor
4/10
Correct use of statistics
Shows awareness of class imbalance and chooses F1-weighted metric; uses classification_report, confusion matrices and ROC-AUC, but lacks statistical tests, uncertainty estimation, multiple-comparison controls and formal significance discussion.
Evidence
Natural-Language-Processing/Text complexity levels/Classification_of_text_complexity_levels.ipynb: selection of 'f1_weighted' in GridSearchCV and classification_report/confusion_matrix usage
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: computation of ROC-AUC, confusion_matrix and printed precision/recall/F1
Data Wrangling & Cleaning
6/10
Preparing and cleaning data
Concrete data-loading and cleaning utilities, encoding fallbacks and regex-based text cleaning are implemented; image paths, label filtering and dataset class for image->time-series conversion are present, showing practical data-wrangling applied to both text and medical-image pipelines.
Evidence
Natural-Language-Processing/Text complexity levels/Classification_of_text_complexity_levels.ipynb: load_subs and clean_subs functions with encoding fallback and regex pipelines
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: dataset filtering (M_Acceptable exclusions) and MedicalDataset.__getitem__/_image_to_time_series implementations
Exploratory Analysis & Visualization
5/10
Exploring and visualizing data
Multiple plotting utilities and diagnostic charts are present (boxplots, confusion matrices, metric barplots, class-mean time-series) and some written interpretation exists, but the narrative is limited and charts are mostly diagnostic rather than deep causal storytelling.
Evidence
Natural-Language-Processing/Text complexity levels/Classification_of_text_complexity_levels.ipynb: boxplot of % words by level and plotted feature-importance
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: plot_time_series_samples, plot_class_comparison and visualization helpers
Predictive Modeling
5/10
Building models that predict
Multiple predictive approaches implemented (Count/Tfidf pipelines with GridSearchCV, CatBoost with text_features, Word2Vec + TensorFlow embedding, and a custom PyTorch 1D CNN), with reasonable evaluation and saving; engineering around training loops and model persistence is present but lacks advanced model debugging, rigorous error analysis or productionized pipelines.
Evidence
Natural-Language-Processing/Classification of Text Complexity W2V and TensorFlow/Vers2_Classification_of_EnglishText_Complexity_Word2Vec_Tensorflow.ipynb: gensim Word2Vec training and TensorFlow embedding model
Natural-Language-Processing/Text complexity levels/Classification_of_text_complexity_levels.ipynb: Pipeline + GridSearchCV over vectorizers and classifiers and CatBoost fit with text_features
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: SpiroCNN class, custom train_model loop and evaluate_model
Business Insight & Impact
3/10
Turning analysis into business value
There is explicit business framing for language-learning customers and an explicit metric choice rationale (precision/recall -> F1), but limited discussion of operational impact, error costs (FP vs FN) in deployment, or integration with product/business metrics.
Evidence
Natural-Language-Processing/Text complexity levels/Classification_of_text_complexity_levels.ipynb: Business task section mentioning foreign language courses and metric selection rationale
SpiroCNN_Gramian-Angular-Field/README.md: description of the medical spirometry classification task and target (M_Acceptable)
Reproducibility & Notebook Hygiene
3/10
Clean, repeatable analysis
Notebooks save models and metadata and set a reproducible STATE variable in places, but there is no pinned environment file, limited seed management, and notebooks contain ad-hoc pip installs and drive.mount steps rather than a reproducible pipeline or CI-ready artifacts.
Evidence
Natural-Language-Processing/Text complexity levels/Classification_of_text_complexity_levels.ipynb: saving model with pickle and STATE variable usage
SpiroCNN_Gramian-Angular-Field/test_demo__SpiroCNN_Gramian_Angular_Field.ipynb: torch.save checkpoint and saving model metadata JSON to Google Drive
Industries
Education• Middle
Health Care• Middle
Technologies
CatBoost
Scikit-learn
Seaborn
Matplotlib
Gensim
spaCy
Pandas
NumPy
NLTK
TF-Keras
Pillow
Torchvision
scikit-image
Recommendations
- Develop NLP classification prototypes for education products (language-level prediction) and harden them into repeatable training pipelines with environment pinning and CI.
- Build small-scale medical imaging POCs (spirometry classification) where the current SpiroCNN and MedicalDataset code can be converted into reusable modules and validated with stronger error analysis.
- Implement reproducibility and MLOps basics - requirements.txt/environment.yml, fixed random seeds, deterministic data splits, and model versioning (DVC or MLFlow).
- Add formal statistical checks and uncertainty quantification for model claims (confidence intervals, bootstrapping, or calibration curves) and explicit FP/FN cost analysis for deployment decisions.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: