class ML_Developer:
    name        = "Mederbek"
    experience  = "24 months"
    focus       = "ML Engineering"
    interest    = ["Agentic AI Engineer", "Research Engineer", "Post-Training & Reasoning", "Alignment & AI Safety"]
    principles  = ["DRY", "KISS", "SOLID"]
    goal        = "AGI Engineer"

    backend = {
        "backend":      ["Python", "FastAPI", "Django", "Django Templates", "DRF", "Django Channels", "Pydantic", "SQLAlchemy", "Alembic", "sqladmin"],
        "platforms":    ["Google Colab", "Kaggle", "Hugging Face", "n8n"],
        "databases":    ["PostgreSQL", "MySQL", "Redis"],
        "devops":       ["Linux", "Docker", "NGINX", "AWS", "Gunicorn", "Uvicorn", "Daphne", "Git", "GitHub", "Postman"],
        "api":          ["REST", "GraphQL", "WebSocket", "gRPC", "SOAP", "HTTP/2", "CORS"],
        "architecture": ["Monolith", "Microservices", "n8n"],
        "libraries":    ["Alembic", "Joblib", "Pillow", "pytest", "Authlib", "passlib", "Streamlit"],
    }

    ml = {
        "ml_ai":              ["PyTorch", "Scikit-learn", "OpenCV", "YOLO", "NumPy", "Pandas", "Matplotlib", "Seaborn", "RoboFlow"],
        "audio_ml":           ["torchaudio", "MelSpectrogram", "AmplitudeToDB", "Resample (сэмплрейт нормализация)", "soundfile"],
        "nlp":                ["torchtext", "LSTM", "BiLSTM", "nn.Embedding", "CountVectorizer", "Naive Bayes (MultinomialNB)", "build_vocab_from_iterator", "HuggingFace datasets"],
        "cnn_architectures":  ["Conv2d/MaxPool2d/AdaptiveAvgPool2d", "BatchNorm2d", "Dropout2d", "VGG-style blocks", "transfer learning patterns"],
        "classic_ml":         ["LogisticRegression", "DecisionTree", "RandomForest", "XGBoost", "SVC", "KNeighborsClassifier"],
        "ml_techniques":      ["class_weight='balanced' (дисбаланс классов)", "stratify (стратифицированное разбиение)", "CosineAnnealingLR / StepLR (scheduler)", "label_smoothing", "collate_fn (кастомный батчинг)", "AdaptiveAvgPool2d (переменная длина входа)"],
        "metrics":            ["accuracy", "F1", "ROC-AUC", "R²", "precision/recall", "classification_report"],
    }

    agent_junior = {
        "llm_basics":    ["OpenAI API / Claude SDK", "Sampling (temperature, top-k, top-p)"],
        "prompting":     ["Few-shot / Chain-of-Thought", "Structured output (JSON mode)", "System prompt design"],
        "agents_core":   [
            "Tool Calling + JSON Schema design",
            "ReAct pattern (Reason+Act)",
            "LangGraph базовый (простые графы, state, edges)",
            "Agent memory (short-term vs long-term)",
            "Multi-agent basics (orchestrator-worker концептуально)",
        ],
        "rag_basics":    [
            "Embeddings концептуально",
            "Qdrant/pgvector — базовый поиск",
            "Chunking стратегии",
            "Hybrid search (BM25 + vector)",
            "Retrieval evaluation (hit rate, базовые метрики)",
        ],
        "context":       ["Sliding window", "Token budget — считать примерно"],
        "eval_debug":    ["LLM-as-judge базово", "Trace logging (LangSmith/Langfuse)"],
        }

    genai = {
        "fine_tuning":   ["LoRA / QLoRA концептуально", "Когда fine-tuning нужен вместо RAG/prompting"],
        "tool_calling":  ["Parallel tool calls", "Forced tool choice / tool_choice параметр"],
        "safety":        ["Guardrails", "Content moderation (входной/выходной фильтр)"],
        "generation_eval": ["Метрики генерации текста (не путать с retrieval-метриками из rag_basics)"],
        }