Bosch, Siemens & Honeywell: AI Engineer

• ANN/MLP: fully connected layers, where every input connects to every neuron. Good for tabular or flat feature vectors. For images it means a huge number of parameters and no spatial awareness.
• CNN: convolutional filters with local connectivity + weight sharing, plus pooling. They capture spatial hierarchies, are translation-equivariant and need far fewer parameters.
• Use CNNs for images, video frames, spectrograms and some time series (1D conv); MLPs for tabular data or as heads on top of embeddings.

• "Imagine standing on a hill in fog, trying to reach the valley. You feel the slope under your feet (the gradient) and take a step downhill (opposite to the gradient). The step size is the learning rate."
• Formally, θ ← θ − η ∇L(θ). Mini-batches estimate the gradient cheaply (SGD); momentum and Adam smooth and adapt the steps.
• Too big a step overshoots or diverges; too small is slow. Use learning-rate schedules and monitor the loss curves.

• Feature extraction: freeze the backbone and train a new head. Use with small data and a similar domain.
• Fine-tuning: unfreeze the top layers (or all) with a low learning rate and discriminative learning rates; augment data.
• For LLMs: prompting / few-shot → RAG for knowledge → LoRA fine-tuning for behaviour or format.
• Domain shift (e.g. factory images vs ImageNet): more unfrozen layers, domain-specific augmentation, possibly self-supervised pre-training on unlabelled factory images.

import pandas as pd
from sklearn.model_selection import train_test_split, cross_val_score
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report
df = pd.read_csv("data.csv"); y = df.pop("target"); X = pd.get_dummies(df, drop_first=True).fillna(df.median(numeric_only=True))
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=.2, stratify=y, random_state=42)
rf = RandomForestClassifier(n_estimators=300, max_depth=None, min_samples_leaf=2, class_weight="balanced", n_jobs=-1, random_state=42)
print(cross_val_score(rf, Xtr, ytr, cv=5, scoring="f1_macro").mean())
rf.fit(Xtr, ytr); print(classification_report(yte, rf.predict(Xte)))

• Mention feature importance, tuning (max_features, depth) and avoiding leakage (fit encoders and imputers on the training data only, ideally in a Pipeline).

• Data: vibration, temperature, current (IoT sensors at high frequency), maintenance logs (failures as labels), operating context.
• Features: rolling statistics, FFT spectral bands, trends; handle sensor drift and missing data.
• Models: anomaly detection (autoencoders, Isolation Forest) where there are few failures; remaining-useful-life regression or survival models once labels exist.
• Edge + cloud: lightweight scoring at the edge for alerts; cloud training and fleet dashboards.
• Value: downtime avoided against false-alarm cost; a technician feedback loop.