AI/ML Engineer

Feature Engineering Techniques

Vì sao: You can design features with target encoding or time-based aggregates without leaking the label, which is what separates a model that holds up in production from one that only looked good offline.

Feature Stores

Vì sao: You can serve the same feature values in training and online inference with point-in-time correctness, removing a common cause of silent training-serving skew.

Supervised & Unsupervised Learning

Vì sao: You can justify a model family and an ensemble choice against the bias-variance trade-off for a given dataset, rather than defaulting to whatever worked last time.

Deep Learning Architectures (CNN/RNN/Transformer)

Vì sao: You can choose between a CNN, a recurrent model or a transformer for the modality at hand, instead of reaching for the newest architecture when a simpler one fits.

LLM API Integration

Vì sao: You can abstract providers behind one interface and manage token budgets, so switching models during a vendor outage or price change is a config change, not a rewrite.

Prompt Engineering for Production

Vì sao: You can version prompts, test them against a fixed set of cases and A/B test a change, so a prompt edit does not regress the feature without anyone noticing.

Model Serving (TorchServe/Triton/vLLM)

Vì sao: You can deploy a model behind Triton, TorchServe or vLLM and tune batching for latency, so the first traffic spike does not turn into timeouts.

Containerized Training (Docker/K8s)

Vì sao: You can package training in a Docker image and run it on Kubernetes, so a colleague can reproduce last week's run on a fresh cluster.

GPU/TPU Resource Management

Vì sao: You can fit a training job into GPU memory with mixed precision and multi-GPU settings, rather than shrinking the model after an out-of-memory failure.

Bias Detection & Mitigation

Vì sao: You can measure a fairness metric across groups before launch and apply a mitigation, so a skewed training set is caught in review and not by affected users.

Model Explainability (SHAP/LIME)

Vì sao: You can use SHAP or LIME to explain one prediction to a business reviewer who asks why a customer was rejected, and back it with global feature importance.