Indie Degree
GitHubSign in
← Programme

Skill graph

75 skills and 93 prerequisites, 8levels deep. Arrows run left to right: everything to the left of a skill has to come first. The programme's study order is value-weighted within these constraints, so this graph is what the ordering is weighed against.

The most depended-on skill is Eval design — 8 skills sit on top of it, more than linear algebra. That is an argument for studying evaluation early that comes from the structure rather than from taste.

has completed evidencenot startedScroll sideways to follow a chain.
Deploying and operatingservicesPython and the numericstackPrompting as engineeringCalculus anddifferentiationLinear algebraProbability andstatisticsTokenisationModel servingReliability patternsPyTorchContext managementFailure modes andhallucinationStructured output andschema enforcementBackpropagation andautodiffOptimisationUnsupervised learningEmbeddings and vectorsemanticsSupervised learningClassical modelevaluationMarkov decisionprocessesToken economicsBatching and throughputLatency budgets andstreamingGPU execution modelPrivacy and datahandlingEval designTool use and functioncallingAttentionNetwork architecturesTraining dynamicsEmbedding retrievalPolicy gradient methodsFeature engineeringValue-based methodsAI economics andbuild-versus-buyPrompt and resultcachingStreaming audiotransportTracing and productionobservabilityChoosing an AI approachBuilding eval datasetsLLM-as-judge and itspathologiesRed-teaming andadversarial testingStatistical significancein evalsPlanning anddecompositionDistributed trainingEvaluating retrievalTransformer architectureVision-language modelsSpeech recognitionSpeech synthesisRegularisation andgeneralisationQuantisation andcompressionAccess-controlledretrievalChunking strategyHybrid and lexicalsearchConversational latencybudgetsRegression testing andCI for modelsTraining data curationCalibrating automatedscores to human judgeme…Implementing atransformer from scratchHuman-in-the-loop designOrchestration patternsDocument understandingIngestion and freshnessRerankingMulti-step reliabilitySandboxing and leastprivilegeSupervised fine-tuningVoice agentsEvaluating agentsParameter-efficientfine-tuningPreference optimisationAlignment techniquesMulti-agent systemsRLHF

Cycles are impossible here by construction — the corpus validator rejects one outright, because a skill that required something which required it back could never be started at all.