Free courses for engineers and students who work with AI, built from what our own teams study.https://kriyaetive.com/https://kriyaetive.com/academy/ai-systems-engineering/from-dartmouth-to-deep-learning/https://kriyaetive.com/academy/ai-systems-engineering/from-dartmouth-to-deep-learning/Two tracks ran for fifty years, knowledge by hand and learning from data. The second won in 2012, and the same pattern decides how you handle a model upgrade today.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/rules-discriminative-generative/https://kriyaetive.com/academy/ai-systems-engineering/rules-discriminative-generative/Rules, discriminative models and generative models each fail differently. Rules break loudly, classifiers decay silently, generated text is plausible but wrong.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/the-transformer/https://kriyaetive.com/academy/ai-systems-engineering/the-transformer/Scaled dot-product attention in one formula, the encoder-decoder and decoder-only shapes, and why append-only generation changes how you order an output schema.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/tokens-embeddings-vector-space/https://kriyaetive.com/academy/ai-systems-engineering/tokens-embeddings-vector-space/What a tokenizer does to your bill, why Tamil costs more tokens than English, and why two embedding models cannot share one index.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/context-windows-and-sampling/https://kriyaetive.com/academy/ai-systems-engineering/context-windows-and-sampling/What shares the context window, why a truncated reply is not an error, how temperature and top-p shape the next token, and why temperature 0 still varies.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/context-limits-and-drift/https://kriyaetive.com/academy/ai-systems-engineering/context-limits-and-drift/Effective context is smaller than advertised, mid-context evidence is used worst, and long conversations drift. A context builder with per-section budgets is the fix.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/latency-and-cost/https://kriyaetive.com/academy/ai-systems-engineering/latency-and-cost/Time to first token, time per output token, the arithmetic of an agent loop that resends its context, and the five levers with what each one costs you.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/from-completion-to-agent-loop/https://kriyaetive.com/academy/ai-systems-engineering/from-completion-to-agent-loop/An agent is a change in who owns control flow. The vocabulary from Anthropic's Building effective agents, and the four limits every loop needs in code.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/react-and-tool-calling/https://kriyaetive.com/academy/ai-systems-engineering/react-and-tool-calling/One turn of the agent loop: the ReAct pattern, what a tool definition guarantees and what it does not, and the runtime rules that keep a tool call safe.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/planning-and-memory/https://kriyaetive.com/academy/ai-systems-engineering/planning-and-memory/Plan-and-Solve, plan-and-execute, and when to replan. Short-term memory is the context window; long-term memory is a database with a write policy.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/multi-agent-orchestration/https://kriyaetive.com/academy/ai-systems-engineering/multi-agent-orchestration/When to split one agent into several, the six-part delegation contract, the single-writer rule for state, and Anthropic's token and performance numbers.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/prompting-rag-fine-tuning/https://kriyaetive.com/academy/ai-systems-engineering/prompting-rag-fine-tuning/The three levers compared, the decision tree that starts from a prompt and an eval set every time, and why fine-tuning is the wrong tool for teaching facts.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/guardrails-and-fallbacks/https://kriyaetive.com/academy/ai-systems-engineering/guardrails-and-fallbacks/Three gates, input, output and action; why privilege has to hold in code a hijacked model cannot change; and a fallback ladder where answers may degrade and actions fail closed.Wed, 16 Sep 2026 00:00:00 GMThttps://kriyaetive.com/academy/ai-systems-engineering/evals-and-observability/https://kriyaetive.com/academy/ai-systems-engineering/evals-and-observability/Offline evals from real traffic, code graders before LLM judges, one trace per request with a span per step, and the loop where every production failure becomes a blocking eval case.Wed, 16 Sep 2026 00:00:00 GMT