Danila Myakotin
Active 8 hours ago
Invite to interview
Message
Download CVCV
Overview
Technical skills
Roles
Overview
A practical ML engineer at an early-to-mid career level who independently built a TF-IDF + MultinomialNB toxic-message classifier and wired it into a VK message bot. The strongest proven skill is building and validating a classical text-classification pipeline with hyperparameter search, evidenced by the Model.ipynb GridSearchCV pipeline and evaluation outputs. The public code lacks production-grade serving, security hardening, experiment tracking, and any novel model research or efficiency engineering.
Technical skills
AI/ML
AI/ML
Scikit-learn
Pandas
NLTK
Jupyter Notebook
Classic ML
Junior AI/ML Engineer
Confidence: Medium ML Engineer
A practical ML engineer at an early-to-mid career level who independently built a TF-IDF + MultinomialNB toxic-message classifier and wired it into a VK message bot. The strongest proven skill is building and validating a classical text-classification pipeline with hyperparameter search, evidenced by the Model.ipynb GridSearchCV pipeline and evaluation outputs. The public code lacks production-grade serving, security hardening, experiment tracking, and any novel model research or efficiency engineering.
Model Architecture & Training
3/10
How well models are designed and trained
Basic model building and hyperparameter search for a classical text classifier (MultinomialNB + TF-IDF) with cross-validation.
Data Pipeline & Feature Engineering
3/10
How data is prepared for models
Simple preprocessing and feature construction - tokenization, lowercasing, removal of URLs and merging labeled data with curated bad/good word lists.
Experimentation & Evaluation
3/10
How results are measured and tested
Standard evaluation using cross-validation and held-out split; prints best_params and computes recall_score, but lacks full experiment tracking or reproducibility artifacts.
MLOps & Deployment
2/10
How models are shipped to production
Minimal deployment: a bot that loads a pickled model and calls VK API to respond to messages; no versioning, monitoring, or secure serving patterns.
Computational Efficiency
2/10
How efficiently computing resources are used
Minor efficiency consideration: parallelized GridSearchCV (n_jobs=-1) used; no GPU/quantization/batching or profiling evidence.
Research Depth & Innovation
1/10
Depth of research and new ideas
No novel research, paper implementations or custom architectures; uses established scikit-learn components without custom layers or algorithmic innovation.
Expertise
Conversational AI & Chatbots• Junior
Industries
Artificial Intelligence• Junior
Internet Services• Junior
Technologies
AI/ML
Recommendations
- Develop prototype content-moderation features and simple chatbots that use classical NLP pipelines and lexicons.
- Improve deployment hygiene by replacing pickle with a safe serialization or model server, and add error handling, logging and rate limiting in the bot.
- Focus on reproducible experiments - add train/val/test artifacts, fixed random seeds, and lightweight experiment tracking (W&B or MLflow).
- Harden preprocessing and tokenization (robust tokenizer, input sanitization) and expand evaluation with confusion matrices and class-wise metrics.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Junior Data Scientist
Confidence: Medium ML Practitioner
A practical ML practitioner at a lower-middle level who can build a simple text-classification pipeline and integrate it into a chat bot. The strongest proven skill is end-to-end supervised NLP model work - evidenced by the TF-IDF + MultinomialNB pipeline, GridSearchCV tuning and the bot integration that loads and uses the model. The work lacks robust evaluation, reproducibility practices, secure model serialization, tests, and clear business-metric reasoning.
Statistical Rigor
2/10
Correct use of statistics
Limited statistical rigor: cross-validation and grid search are used, but there are no assumption checks, uncertainty estimates, calibration, or careful metric usage; a recall_score call appears to use arguments in the non-standard order.
Evidence
MLBot/Model.ipynb: use of GridSearchCV(cv=5, scoring='f1_weighted') and param_grid definitions
MLBot/Model.ipynb: recall_score(nb_tfidf.predict(X_test), y_test) (metric usage)
Data Wrangling & Cleaning
3/10
Preparing and cleaning data
Basic data wrangling is present - lowercasing, simple URL filtering, custom tokenizer and combining labeled data with word-lists - but dataset merging and the train/test split are done by naive slicing which risks leakage.
Evidence
MLBot/Model.ipynb: df['comment'] = df['comment'].str.lower() and df[~df.comment.str.contains("http")]
MLBot/Model.ipynb: def tokenizer(...) and X = np.concatenate((df['comment'].values, bad_words[0].values), axis=0) with slicing for X_train/X_test
Exploratory Analysis & Visualization
1/10
Exploring and visualizing data
Exploratory analysis is minimal - there are a few summary prints but no visualizations, structured EDA, or written interpretations of findings.
Evidence
MLBot/Model.ipynb: sum(df['toxic'].values) and print outputs without interpretive cells
Predictive Modeling
3/10
Building models that predict
Predictive modeling shows correct use of TF-IDF, MultinomialNB and GridSearchCV with parallel CV and predict_proba-based decision logic, but lacks calibration, error analysis, robust CV design, and proper baseline/control experiments.
Evidence
MLBot/Model.ipynb: Pipeline([('vect', TfidfVectorizer()), ('clf', MultinomialNB())]) and GridSearchCV setup
MLBot/Model.ipynb: use of nb_tfidf.predict_proba([...]) and thresholding, and later pickle.dump of the fitted GridSearchCV
Business Insight & Impact
1/10
Turning analysis into business value
There is a clear application goal (detect toxic messages) and a deployed action (bot sends a warning), but no business-metric framing, cost-of-error analysis, or FP/FN tradeoff discussion.
Evidence
MLBot/README.md: describes a bot that detects toxic messages
MLBot/BotForVK.py: Send(...) and logic to message users when prediction indicates toxicity
Reproducibility & Notebook Hygiene
1/10
Clean, repeatable analysis
Low reproducibility and hygiene: model artifacts are pickled, there is no pinned environment or requirements, no fixed random seeds, and the notebook has negligible explanatory reasoning.
Evidence
MLBot/Model.ipynb: pickle.dump(nb_tfidf, open(os.path.join('model_new.pkl'), 'wb'))
MLBot/Model.ipynb: '# NOTEBOOK META: cells=24, reasoning_ratio=0, exec=iterative, outputs=1'
Industries
Internet Services• Junior
Technologies
Classic ML
Jupyter Notebook
Scikit-learn
Pandas
NLTK
Recommendations
- Replace pickle-based model exchange with a safer, versioned serialization or model-serving approach and add validation around model loading to avoid RCE risks.
- Adopt proper train/test splitting (or stratified/time-based splits as needed), set random seeds, and add calibration plus error-analysis (confusion matrix, per-class metrics and threshold tuning) to quantify uncertainty and FP/FN tradeoffs.
- Add reproducibility artifacts: requirements.txt or environment.yml, notebook narrative cells describing decisions, and small unit/integration tests for tokenization and the bot message flow.
- Improve data hygiene and EDA: add visualizations, investigate class balance and label quality, and document data provenance and preprocessing choices to reduce leakage risk.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: