MODULE 02 · FDE ACADEMY
Build
This module is the engineering core. It assumes you can use a terminal and read documentation, and it does not assume any machine learning background — none of the work below involves training a model.
- Lessons
- 5
- Module
- 02 of 5
Every lesson is built around a recipe on this site, because a recipe is a set of steps grounded in the project’s own documentation, with the page it was copied from attached. Read the lesson for the reasoning, then run the steps and read the evidence behind them.
The order matters. Retrieval quality is bounded by what you understood in the first two lessons, and the last two lessons are what turn a working laptop demo into something that survives a reboot.
- Covers
- APIsLLM fundamentalsRAGworkflow automationdatabasesDocker
Free and self-paced. No certification is offered.
Lessons
5 lessons
Each one ends with a task carried out somewhere on this site.
- 01What a language model does, and what it costs to runTokens, context windows and quantisation at working depth, then the memory arithmetic that decides which GPU you need — weights, KV cache and overhead, each with its inputs on the page.About an hour, with a calculator open.llm-inference5 steps
- 02The API is an HTTP endpoint you can curlChat completions, streaming, structured output and the four failure modes worth designing for. The same interface whether the model is in the next rack or in someone else’s cloud.Around ninety minutes, most of it waiting for a model to load.api-integrationllm-inference5 steps
- 03Retrieval is the productChunking, embeddings, hybrid search and re-ranking — and the discipline that separates a working deployment from a demo: evaluating retrieval on real questions before anyone is invited in.A long afternoon: the deployment takes an hour, the evaluation takes the rest.ragllm-evaluationvector-databases5 steps
- 04The database is where it becomes a systemOne PostgreSQL with a vector extension instead of two data stores, why the index is a recall decision rather than a speed decision, and the delete you must be able to run.About two hours against a deployment you already have running.postgresqlvector-databasesdata-engineering5 steps
- 05Containers, and the wiring between themImages, volumes, networks and one compose file — then the automation layer where the integrations live, which is the part that connects a working stack to the work people actually do.Two to three hours, including a deliberate reboot.dockerworkflow-automationlinux5 steps