Learn AI engineering,
in the right order.

What machine learning is, training vs inference, neural nets, embeddings, and why GPUs. The base layer.

Tokenization, attention, how an LLM generates text, context windows, and why models hallucinate.

The layer underneath every LLM, from the byte up: bytes, Unicode and UTF-8, byte-level BPE training, special tokens, encode/decode, and performance.

Steering a model for real, structured output, and treating context as a budget.

Retrieval pipelines, chunking, embeddings, vector search, reranking, citations, and where RAG quietly breaks.

Tool use, the agent loop, what a harness owns, orchestration, failure modes, and multi-agent patterns.

Task evals, LLM-as-judge, tracing, regression testing, and production monitoring for AI features.

Threat modeling, prompt injection, sandboxing, least-privilege tools, data leakage, output guardrails, and defense-in-depth harness design.

Datasets, labeling, synthetic data, deduplication, contamination, and why data quality dominates model behavior.

When to fine-tune, dataset design, LoRA/PEFT/QLoRA, evals, catastrophic forgetting, and deployment.

Preference datasets, reward models, RLHF, DPO, GRPO, evals, and when RL on open models is worth it.

Model distillation, teacher-student data, quantization, pruning, evals, and deployment tradeoffs.

Prefill vs decode, the KV cache, memory bandwidth, context length, output tokens, and the numbers that explain latency.

Batching, continuous batching, speculative decoding, streaming, and the scheduler that ties them together.

vLLM/TGI/llama.cpp, GPU cost per token, API vs self-hosting, edge, and production operations.

Sound waves, sample rate, bit depth, channels, PCM, codecs, FFT, spectrograms, and mel features for voice AI.

Finding speech in audio: energy thresholds, denoising, WebRTC VAD, neural VAD, streaming latency, and production evaluation.