Machine Learning Training in Pune with Placement

A targeted refresher rather than an introduction — vectorised NumPy 2 operations, pandas 2 with Arrow backend (10–100× faster than legacy pandas on common joins and groupbys), idiomatic Python with type hints, and the discipline of writing reproducible Jupyter notebooks. We assume Python fluency on day 1; this week levels the floor on data-engineering vocabulary that the rest of the course depends on. By the end of week 1 every student has a clean development environment with uv / poetry, ruff, mypy, and Jupyter Lab.

The math that hiring panels actually test for and that you will use weekly — not full graduate-level proofs. Linear algebra (vectors, matrices, eigenvalues, SVD with intuition), calculus for ML (gradients, partial derivatives, the chain rule applied to backpropagation), probability and statistics (distributions, MLE, Bayesian basics), and information theory primitives (entropy, cross-entropy, KL divergence). Worked through Python rather than chalkboard — every concept illustrated by code that you run.

scikit-learn 1.5+ as a production library, not a tutorial toy. Linear and logistic regression with regularisation (L1, L2, ElasticNet), tree-based models (CART, Random Forest, gradient boosting). Boosting libraries — XGBoost, LightGBM, CatBoost — and the honest comparison of when each is best. Support Vector Machines, k-NN, naive Bayes for completeness. Throughout the module we drill the discipline that prevents the leak-and-overfit failure mode that ends most fresher ML interviews — proper cross-validation, leak-free pipelines via Pipeline + ColumnTransformer, target leakage detection, and cross-validation strategies for time-series, grouped, and stratified data.

Clustering (k-means, hierarchical, DBSCAN, HDBSCAN), dimensionality reduction (PCA, t-SNE, UMAP), and density estimation. Then the higher-leverage skill — feature engineering — where most production-model performance gains come from. Numerical encoding choices, categorical encoding (one-hot, target, frequency, ordered target), date / time features, text features (TF-IDF, hashing trick), interaction features, and the Featuretools / tsfresh libraries for automated feature engineering on tabular and time-series data.

Fraud, churn, defaults, defect detection — the problems businesses actually pay for are almost all imbalanced, and accuracy is a useless metric on all of them. You work through resampling (SMOTE and its variants, undersampling, and why naive oversampling before the split leaks), class weighting, and threshold optimisation driven by the cost of each error type rather than a default 0.5. Evaluation shifts to precision-recall curves, PR AUC and calibration. The framing throughout is business cost: a false negative and a false positive rarely cost the same, and the model should reflect that.

Unsupervised work is where practitioners most often produce confident nonsense, because there is no test-set score to catch them. K-Means with the elbow and silhouette methods and its assumptions about cluster shape; hierarchical clustering and dendrogram reading; DBSCAN for arbitrary shapes and noise; Gaussian mixtures for soft assignment. Then dimensionality reduction — PCA and what the components actually mean, t-SNE and UMAP for visualisation, and the standing warning that both will happily produce beautiful, meaningless structure if you tune them until it appears.

Ensembles win most tabular problems, and most competitions, so they get treated properly rather than as a footnote on decision trees. Bagging and why averaging decorrelated trees reduces variance; Random Forest and its out-of-bag estimate; then boosting as a fundamentally different idea — fitting sequentially to the residual. XGBoost, LightGBM and CatBoost are covered as distinct tools with different strengths (sparse data, leaf-wise growth, native categorical handling) rather than interchangeable names, along with the hyperparameters that actually move the needle and the ones that do not.

PyTorch 2.4+ as the dominant framework for new deep-learning work. Build neural networks from scratch (forward / backward pass, autograd, optimisers, schedulers), then move to convolutional networks for image classification (transfer learning with ResNet / EfficientNet / ViT), and recurrent / transformer architectures for sequence work. Cover regularisation (dropout, batch norm, layer norm, weight decay), training stability (gradient clipping, mixed-precision, learning-rate warmup), and the discipline of debugging a deep network when training stalls — which it will.

Hugging Face Transformers as the default library for using pre-trained models. Fine-tune a BERT / RoBERTa model for classification, build sentence embeddings for semantic search, do named-entity recognition with token-classification heads, and on the vision side fine-tune a ViT or EfficientNet for image classification. Cover the Hugging Face training stack (Trainer, datasets, tokenizers, accelerate) and the inference stack (pipelines, optimum, ONNX export).

The 2026 differentiator on Pune ML Engineer resumes. Cover parameter-efficient fine-tuning (LoRA, QLoRA, PEFT) on open-source models (Llama 3.x, Mistral, Phi-3) running on a single consumer GPU or Colab Pro. Then RAG — chunking strategies, embedding models, vector databases (pgvector, Chroma, Weaviate), retrieval techniques (BM25 + dense hybrid, reranking with cross-encoders), and a small production-style RAG service. Includes evaluation — RAGAS, retrieval recall, end-to-end answer correctness — because LLM-augmented systems without evaluation are theatre, not engineering.

The week that separates ML Engineer titles from Data Scientist titles. MLflow for experiment tracking, model registry, and model versioning. FastAPI for inference endpoints (sync and async, with Pydantic v2 schema validation, JWT auth, structured logging). Docker for inference containers, GitHub Actions CI/CD pipelines, basic monitoring (Prometheus + Grafana), and an honest discussion of model drift, data drift, and the human cost of silent model failures in production.

Two weeks of full-time capstone work plus structured interview preparation. Pick one of three capstone projects (see Capstone Projects). Mock interviews calibrated for Pune ML hiring panels — Tiger Analytics, Fractal, Persistent, BMC, BMW TechWorks, Mercedes-Benz R&D. Includes an algorithm round (questions on bias-variance, overfitting, regularisation, evaluation metrics — the 30-question core that every Pune panel asks), a coding round (NumPy / pandas / scikit-learn live coding), and a system-design round on serving a model at scale. Resume / LinkedIn / GitHub polish included.