DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///
DARC.DEV

[ 02 — Developer · 6 lessons · ~4.2h with exercises ]

Build with LLMs

Turn model responses into reliable software. Build validation, retrieval, evaluations, observability and cost accounting around one small application.

← All learning paths

Your progress

0/6

Mark completion after doing the exercise. The checklist records your self-assessment, not a certificate. Stored for this browser only.

Module 01

Contracts before cleverness

Build around the failures your app must handle.

~40 min · read + exercise

Structured output and domain checks

Distinguish valid JSON from correct data.

Exercise

Design an invoice extraction object with missing fields and review states.

Done when: A missing due date does not become a guessed due date.

Read the lesson

~35 min · read + exercise

Design the context policy

Define source authority, freshness and budgets.

Exercise

Write a context policy for a support assistant using two conflicting documents.

Done when: The policy explains which source wins and when the answer must abstain.

Read the lesson

Module 02

Measure the pipeline

Observe why a result is good or bad.

~50 min · read + exercise

Hybrid retrieval and permission filters

Evaluate evidence retrieval before generation.

Exercise

Label ten queries including exact identifiers, paraphrases and cross-user access cases.

Done when: The expected supporting passage and permission scope are specified for every query.

Read the lesson

~40 min · read + exercise

Useful traces without raw secrets

Design redacted spans and outcome metadata.

Exercise

Create a trace envelope for retrieval, model, validation and write steps.

Done when: A reviewer can locate a failed stage without receiving the customer's raw secret.

Read the lesson

Module 03

Release with confidence

Connect evaluation, billing and rollback.

~50 min · read + exercise

An evaluation gate in CI

Make a known regression fail the release.

Exercise

Define deterministic assertions and deliberately fail one test in your runner.

Done when: The failing run has a non-success outcome and an inspectable artifact.

Read the lesson

~35 min · read + exercise

Compute the full inference bill

Account for uncached input, writes, reads and output.

Exercise

Reproduce the worked prefix example and change the hit rate.

Done when: You can explain why total task savings differ from prefix savings.

Read the lesson