01Learning path
Build user-facing applications powered by LLMs, from the first API call to a deployed, evaluated service.
input · support email
“Hi — I was charged twice for the annual plan on my team account this morning. Can someone refund the duplicate? Order 48213.”
output · TicketTriage
{
"category": "billing",
"priority": "high",
"order_id": "48213",
"needs_human": true
}
01
This path is for software engineers who want to add LLM-powered features to real products. It starts with the fundamentals that shape every design decision — tokens, context windows, model APIs, and structured outputs — and builds toward complete applications that combine retrieval, tool calling, and a proper API layer.
Throughout the path you treat LLM features like any other production code: typed inputs and outputs, tests, versioned prompts, and clear error handling. Python is used for back-end services, with TypeScript examples for the front-end pieces that stream responses to users.
You finish with a deployed application, an evaluation suite that guards it against regressions, and the written documentation to explain how it works and where it falls short.
Who it’s forBackend and full-stack developers who want to design, build, and ship LLM features inside real applications.
02
03
BuildWrite a small Python client wrapper that logs token usage and handles retries.
MilestoneYou can predict a request's token cost and explain what happens when the context window is exceeded.
BuildBuild a support-ticket classifier that returns validated category, priority, and summary fields.
MilestoneYour extraction function returns schema-valid output on every case in a small test set.
BuildAdd document retrieval and one tool call to your application, returning answers with citations.
MilestoneAnswers include citations that point to the passages they were drawn from.
BuildExpose the application through a FastAPI service and a minimal streaming web client.
MilestoneA user can ask a question in the browser and watch a cited answer stream in.
BuildContainerize the service, add an evaluation suite to CI, and review traces from test traffic.
MilestoneYour application runs in a container and a failing evaluation blocks a merge.
04
05
output · TicketTriage
{
"category": "billing",
"priority": "high",
"order_id": "48213",
"needs_human": true
}
Build a service that classifies incoming support tickets and extracts key fields into a validated JSON schema, ready for routing and reporting.
Build an assistant that answers questions over a private document set, retrieves the passages it relies on and cites them in every answer. You own the full pipeline, from ingestion to source-aware responses.
Instrument an LLM application with tracing, then build dashboards and alerts for cost, latency, errors and output quality signals.
06
Outcomes describe skills and portfolio artifacts. No employment outcome is promised.
Build the foundations, projects, and habits needed to design more dependable AI applications.