One path, two tracks. Product is the GenAI engineering — design, build, tune, serve, and operate LLM systems end to end. Platform is the cloud they run on — accounts to GPUs to pipelines. Eighteen stages, every concept drawn, every number worked by hand.
Product rides Lines A→C — model, build, operate. Platform is Line D — the cloud underneath, taught through a GenAI lens. Stage 08 is where platform meets product. Click any station to board.
Never touched AI before? Watch this first. Plain-language videos that build the intuition every stage on this map assumes. No math, no code, no prerequisites.
The product is delivered by GenAI. The platform is delivered by cloud. Browse either — every Platform session is taught through a GenAI lens, so nothing here is platform trivia.
What a language model actually is — tokens, embeddings, attention, sampling — capped by building a tiny GPT from scratch.
Open →When to tune and when not to, LoRA & QLoRA mechanics against a 24 GB ceiling, data curation, then compressing and serving the result.
Open →From imitation to preference: reward models, PPO & GRPO, DPO and its variants, reasoning & test-time compute, and the evaluation gate.
Open →The open-book exam: ingestion, indexing, hybrid retrieval & reranking, agentic and graph RAG, evaluation, and the production wall.
Open →The full serving stack — hardware, batching, KV-cache, quantization, speculative decoding — and the optimizations that pay for the GPUs.
Open →FastAPI contracts, async concurrency, token streaming from vLLM to humans, and the protect-and-optimize gauntlet before launch.
Open →Evaluation as a decision system: the two clocks (offline & online), tracing a token machine, and calibrating LLM judges you can trust.
Open →Model weights as data, GPUs on Kubernetes, and production serving topologies — the station where platform meets product.
Open →Regions, zones, accounts, and the resource vocabulary — the skeleton every later stage hangs services on.
Open →Who can do what, where: principals, policies, and the guardrails your model endpoints and data inherit.
Open →VMs to serverless to GPU fleets — instance families, quotas, spot economics, and what inference actually rents.
Open →VPCs, private endpoints, load balancers, and egress — including what moving model weights really costs.
Open →Hot-to-archive tiers, object stores for weights and vector data, and the durability & cost math behind them.
Open →Sync vs async, queues, events, and the coupling decisions behind GenAI services that survive load.
Open →SLOs, DR tiers from backup to multi-site, and observability for systems that must not drop tokens.
Open →Preventive & detective controls, budgets and tagging, and where GPU spend hides — FinOps with an LLM bill.
Open →The R-spectrum from rehost to refactor — moving workloads (and models) without breaking them.
Open →Infrastructure as Code and pipelines — plan, apply, destroy, and delivery flows that ship model services.
Open →Free forever. Built by engineers who ship this stack. When you finish, you won't hold a certificate — you'll hold the skills to deliver GenAI products and run the platform beneath them.