Learn the mechanism. Keep the evidence.
InterviewsVector’s original AI Engineering Academy connects derivation, code, failure analysis, production constraints, and interview defense.
Open the capability roadmap- Arcs
- 11
- Labs live
- 76
- Mapped
- 76
Published original labs
01Draw the AI System BoundarySeparate deterministic software, probabilistic model behavior, data, and human judgment before choosing tools.55 min →02Reproducible Experiments, Not Reproducible NotebooksMake code, data, seeds, dependencies, and hardware assumptions inspectable.80 min →03Data Contracts for ModelsSpecify schema, meaning, time, lineage, and allowed use before training or retrieval.70 min →04Define the Measurement Before the ModelTranslate a product claim into offline, online, safety, latency, and cost measures.65 min →05A First AI Workload Cost ModelEstimate tokens, accelerator time, storage, and network movement before committing architecture.75 min →06From Experiment to Service BoundaryExtract a tested model contract from exploratory code without copying notebook state into production.75 min →07Vectors as RepresentationsConnect coordinates, basis, scale, and information loss to learned representations.80 min →08Similarity Is a Retrieval PolicyDerive dot product and cosine similarity, then show why normalization changes what a retrieval system rewards.95 min →09Matrices as Small ProgramsRead a matrix as a composition of projection, scaling, rotation, and mixing operations.90 min →010Probability for Decisions, Not DecorationUse conditional probability, calibration, and expected cost to make model outputs actionable.95 min →011Optimization Under NoiseDerive stochastic gradients and distinguish noisy progress from broken learning.100 min →012Numerical Stability Is Part of the AlgorithmBreak softmax, log likelihood, and low-precision reductions, then repair them.100 min →013Complexity for Tensor ProgramsEstimate FLOPs, memory traffic, activation storage, and asymptotic traps from tensor shapes.85 min →014The Generalization ContractDefine the population, decision, loss, and evidence under which performance is expected to transfer.90 min →015Split Data by Causality, Not ConvenienceUse information availability, time, entity, group, and provenance to construct honest evaluation partitions.105 min →016Linear Models as Debugging InstrumentsUse transparent linear baselines to expose target, feature, split, scale, and slice failures before adding complexity.110 min →017Trees, Boosting, and the Shape of Residual ErrorUnderstand how tree ensembles partition mistakes and when their inductive bias wins.110 min →018Calibration, Thresholds, and Decision CostTurn scores into actions using reliability curves and asymmetric error costs.100 min →019When the Product Needs Ranking, Not ClassificationChoose ranking losses and metrics when ordering quality matters more than labels.90 min →020Drift, Feedback Loops, and Delayed LabelsDesign monitoring when the system changes the data it later learns from.110 min →021A Neuron Is a Parameterized Decision SurfaceConnect affine maps and nonlinearities to representation capacity.80 min →022Computation Graphs Make Learning InspectableRepresent forward values, dependencies, and local derivatives explicitly.90 min →023Backpropagation as Local ContractsDerive reverse-mode autodiff as small, testable vector-Jacobian products rather than one mysterious global formula.120 min →024Read Optimization DynamicsUse loss curves, gradient statistics, and parameter updates to separate data, optimization, and capacity failures.105 min →025Normalization and Residual PathsUnderstand why signal and gradient paths shape trainability in deep networks.105 min →026Convolutions and Locality as an Inductive BiasBuild a convolution and connect weight sharing to images, audio, and structured grids.105 min →027Neural Training Incident: Loss Becomes NaNTrace a realistic failure through inputs, precision, activations, gradients, and optimizer state.110 min →028Tokenization as a Compression ContractStudy how vocabulary construction changes sequence length, multilingual behavior, cost, and failure modes.105 min →029Derive Attention from Content-Based RoutingBuild scaled dot-product attention from the need to route information between positions.120 min →030Multi-Head Attention Is Parallel Representation RoutingExplain head dimension, projections, and what head diversity does and does not guarantee.95 min →031Position, Context, and ExtrapolationCompare positional mechanisms by the invariants they encode and how they fail outside training lengths.105 min →032Assemble and Test a Transformer BlockCompose attention, MLP, normalization, masking, and residual paths with shape tests.140 min →033Pretraining Objectives Shape Model BehaviorConnect next-token prediction, masking, data mixtures, and preference objectives to observable capabilities.105 min →034Decoding Is a Product PolicyTreat temperature, top-p, constraints, and stopping as explicit quality and risk decisions.90 min →035The KV Cache Capacity PlanDerive per-token cache memory, then connect context length, concurrency, precision, batching, and paging to serving capacity.120 min →036Prompt, Retrieve, Fine-Tune, or Train?Choose the smallest intervention that changes the missing capability or knowledge.95 min →037Supervised Fine-Tuning as Behavior ShapingDesign data, masking, batching, and evals for instruction tuning without confusing memorization for improvement.125 min →038LoRA and the Low-Rank Update HypothesisDerive parameter and memory savings, then test where low-rank adaptation stops being sufficient.120 min →039Preference Optimization and Reward FailureCompare preference objectives through their data assumptions and reward-hacking risks.115 min →040The Vision-Language BridgeTrace images through patching, encoders, projectors, fusion, and autoregressive decoding.120 min →041Multimodal Production BoundariesEngineer media ingestion, latency, privacy, safety, and modality-specific evaluation.105 min →042The Embedding ContractVersion task, model, preprocessing, distance, and dimensionality as one inseparable interface.90 min →043Approximate Nearest Neighbors Under a Latency BudgetCompare graph and partitioned indexes through recall, build cost, memory, filtering, and updates.115 min →044Chunking Is a Recall PolicyConnect document structure and answer span length to retrieval candidates and context waste.95 min →045Hybrid Retrieval and Score FusionCombine lexical and semantic evidence without comparing incompatible raw scores.110 min →046Reranking and Context AssemblySeparate candidate generation, expensive relevance judgment, diversity, and final context packing.110 min →047Diagnose RAG by StageMeasure corpus, retrieval, reranking, context, generation, citation, latency, and cost separately so failures remain attributable.125 min →048Retrieval Security, Deletion, and FreshnessPreserve tenant ACLs, deletion semantics, provenance, and index freshness through ingestion and ranking.105 min →049Tools Are Typed Authority BoundariesDesign tool names, schemas, errors, idempotency, and authorization for safe model use.105 min →050A Tool-Using Agent Is a Bounded State MachineImplement budgets, typed transitions, approval gates, idempotency, and terminal states around a probabilistic planner.135 min →051State, Memory, and Context Are Different ThingsSeparate durable workflow state, retrievable memory, and model-visible context.95 min →052MCP Host, Client, and Server BoundariesImplement protocol capabilities without delegating consent, authorization, or validation to the model.115 min →053Durable Agent WorkflowsSurvive retries, restarts, partial tool success, human pauses, and changing external state.120 min →054When Multiple Agents Are Actually JustifiedUse ownership, isolation, parallelism, and context constraints—not personas—to decide decomposition.100 min →055Agent Red-Team: Untrusted Instructions Meet Real ToolsTest prompt injection, confused deputy behavior, data exfiltration, and excessive agency.120 min →056An Eval Is a Decision SystemStart from a release decision, then choose cases, graders, uncertainty, and escalation.105 min →057Build a Golden Dataset That Can Disagree with YouSample real tasks, hard negatives, long tails, and policy boundaries without teaching to the test.110 min →058Calibrate LLM JudgesMeasure position bias, verbosity bias, self-preference, variance, and disagreement against human labels.115 min →059Online Evaluation Without Shipping BlindConnect shadowing, canaries, holdbacks, user signals, and rollback thresholds.105 min →060Observe the Decision Path, Not Just the Model CallTrace retrieval, prompts, tools, model versions, policy checks, latency, and cost with controlled data capture.110 min →061Build a Safety Case for an AI FeatureTie hazards to controls, evidence, residual risk, ownership, and launch conditions.105 min →062AI Incident Response and Correction LoopsContain harm, preserve evidence, distinguish model and system causes, and turn findings into durable gates.110 min →063The Accelerator Memory HierarchyReason from registers, SRAM, HBM, interconnect, and host memory to wall-clock performance.110 min →064Precision and Quantization as Error BudgetsCompare numeric formats and quantization methods through memory, throughput, calibration, and quality loss.120 min →065Continuous Batching and QueueingModel prefill, decode, scheduling, tail latency, and throughput under mixed request shapes.120 min →066Parallelism for InferenceChoose tensor, pipeline, expert, and data parallelism from model shape and interconnect constraints.120 min →067Distributed Training Without Topology AmnesiaMap optimizer state, gradients, activations, parameters, and collectives onto an actual cluster.135 min →068The Model-Serving Control PlaneDesign placement, rollout, autoscaling, health, model loading, and rollback independent of request execution.115 min →069Load-Test an AI ServiceGenerate representative sequence shapes, concurrency, cancellations, and cache behavior while protecting quality signals.110 min →070Design a Multi-Tenant AI GatewayCreate one policy and observability boundary for provider routing, tenant isolation, budgets, fallback, and release evidence.145 min →071Build, Buy, or Partner for AI CapabilityCompare control, differentiation, switching cost, data rights, reliability, and total operating cost.105 min →072Draw the AI Platform BoundariesDecide what the platform standardizes and what product teams must continue to own.110 min →073The Cost–Quality–Latency Operating EnvelopeUse routing, caching, batching, model choice, and graceful degradation within explicit product constraints.115 min →074Governance as a Control PlaneEncode inventory, ownership, evidence, policy, approvals, and exceptions into shipping workflows.110 min →075Migrate a Production AI Stack Without a Flag DayDual-run, shadow, compare, canary, reconcile, and retire old paths with rollback intact.120 min →076AI Architect Capstone: Defend the SystemProduce and defend a complete architecture under changing scale, safety, cost, and organizational constraints.180 min →