The Seven Dimensions of Applied AI

Model selection is 1% of the design work in a production applied AI system. The other 99% happens across six dimensions that most applied AI coverage collapses into vague "infrastructure" talk or skips entirely.

The framework is called The Seven Dimensions of Applied AI. The dimensions are Data Architecture, Domain Logic, Orchestration, Models & Agents, Governance & Containment, Human Judgment, and Calibration.

This is what each dimension does, and why it matters more than the model running inside it.

Data Architecture Is the Decision About What the System Can See

Three concerns live here: signal design, data readiness, and data absence. Signal design is choosing what to measure and why. Data readiness is assessing whether a specific measurement is trustworthy before inference starts. Data absence is surfacing what's missing before any human reviews the output.

A pick-and-place robot operating on stale positioning data moves with full confidence toward coordinates that no longer reflect the floor. The collision shows up in the incident report, not the model logs. That failure mode is identical whether the system routes robots or scores creators: a dimension upstream gets skipped, and every downstream component inherits the error without knowing it.

I am building TikSense, an agent-native multi-model creator vetting decision engine for TikTok Shop sellers, built on The Seven Dimensions of Applied AI. TikSense pulls from three parallel data sources per creator. For a meaningful portion of creators, one of those sources returns nothing. The system flags that gap before multi-model analysis begins. A model receiving incomplete input doesn't know it's incomplete. Data Architecture is what does.

Data Architecture sets the ceiling for everything else. No downstream dimension compensates for decisions made poorly here.

Domain Logic Is the Encoded Point of View

Models arrive with general capability and no position on your specific problem. Domain Logic supplies that position through deterministic rules written by someone who understands the problem at depth, applied before any model sees anything.

Platforms like Douyin illustrate this: signals optimized for watch time and signals optimized for purchase intent routinely diverge for the same creator. A creator who holds attention produces different signals from a creator who drives transactions. The Domain Logic layer decides which signal wins based on context, before any model does anything with the result. That classification logic doesn't emerge from training data. It comes from someone who watched the difference between attention and conversion play out at scale and encoded that distinction as explicit logic.

Building TikSense, I encountered this directly. Standard engagement metrics collapsed to near-zero for TikTok creators regardless of actual selling power. A creator doing consistent commerce volume scored 49 on a standard engagement algorithm. The metric was correct and useless for evaluating commerce creators. The fix was detecting creator type before scoring and switching to a commerce-specific approach before passing anything upstream. That detection and switching logic is Domain Logic.

Orchestration Is How Dimensions Communicate, in What Sequence, and With What Reasoning

A ride-hailing platform routing 10 rides a day can survive on informal coordination. At 10 million rides a day, pricing agents, dispatch agents, route agents, and ETA agents negotiate through orchestration logic in under 800 milliseconds to produce one coherent output. Volume reveals what structure the system actually requires. At production scale, the orchestration layer isn't optional. It's the only mechanism capable of coordinating components that each hold a partial picture of the problem.

In TikSense, a routing agent runs before four parallel analysis tracks fire. The routing agent reads creator signals and shapes everything downstream: which scoring logic applies, which model prompts activate, which data sources receive priority. Removing that step makes the pipeline faster and every analysis less accurate. I kept it. The cost is the point.

Agent-native architecture is a maturation of this dimension. Orchestration components that gain the ability to reason, assess, and route within bounded authority turn a static pipeline into a system capable of handling what it wasn't designed to expect.

Models and Agents Handle What Structured Logic Alone Can't, and Are the Most Substitutable Parts of the System

A 2026 review of autonomous vehicle perception research confirms the counterintuitive finding: adding more models with narrower scope produces more reliable output than using fewer models with broader scope. Specialized models for vehicles, pedestrians, cyclists, and road signs run in parallel. A fusion agent reconciles their outputs. A vision model that excels at vehicle detection may struggle with cyclist pose estimation. Running both in parallel and reconciling the results covers failure modes that no single model handles across all conditions. Specialization plus reconciliation outperforms generalization.

TikSense runs multi-model, multi-agent analysis across four parallel tracks: content classification, audience assessment, commerce signal detection, brand alignment. Model selection for each track took an afternoon. Designing the agent architecture, defining each agent's authority boundaries, and verifying that outputs from different models align into a coherent assessment took months.

The benchmark conversation in AI conflates model capability with system performance. A well-scoped model inside a well-designed multi-agent system outperforms a frontier model applied to an unbounded problem without supporting architecture. The other six dimensions are the product. Models and agents are the infrastructure inside it.

Governance and Containment Is the System's Rules About Itself, Defined Before Anything Runs

This dimension answers three questions: what can each agent touch? When does the system stop and surface a decision to a human? How do disagreements between agents get handled? These constraints are built into the system's architecture, not applied as a policy layer after deployment.

Multi-agent systems require hierarchical authority structures. A classification agent classifies but doesn't recommend. An analysis agent recommends but doesn't execute. An execution agent acts within defined parameters but escalates above a threshold. Each agent operates within a defined authority tier. Authority determines what an agent is allowed to decide, not what it is capable of performing. I call this the Agent Authority Escalation Model.

In TikSense, when Domain Logic and the AI model disagree on a creator assessment, the system surfaces a flag. It doesn't merge the two into an average. It doesn't silently pick one. The disagreement itself is the output, and a human reviews it.

Alignment and containment are not the same thing. A model can internalize values and still find the gap between what an instruction covers and what the environment permits. Containment designed for cooperation is a different architecture from containment designed for capability. Article 9 of the EU AI Act, Regulation (EU) 2024/1689, mandates risk management systems as a technical requirement for high-risk AI, auditable by construction rather than reconstructable after an incident, which is the only moment it can be done correctly.

Human Judgment Enters When the System Reaches the Boundary of Its Own Confidence

Human Judgment in applied AI gets framed as a safety net for when AI fails. That framing produces the wrong architecture. By the time a system produces an observable error, the architectural decision that caused it was made earlier.

Consumer platforms processing millions of content decisions a day make this asymmetry visible. Moderation agents do the volume work that no human team could cover. One wrong call on a high-profile account generates a news cycle. The agents identify. Humans act on the calls that exceed agent authority. That boundary is invisible when it works and expensive when it doesn't.

In TikSense, when the multi-agent analysis produces a disagreement between two assessment tracks, the system presents both outputs and the reasoning behind each. The reviewer sees the disagreement, not a blended score. That reviewer brings context no agent carries: knowledge of the seller's specific product line, past experience with similar creator partnerships, commercial intuition about category fit.

Human Judgment also refines the other six dimensions over time. Reviewers' decisions update what signals Data Architecture tracks, which Domain Logic rules need adjustment, which agent authority boundaries need tightening, and which Governance & Containment thresholds should move.

Calibration Is How the System Improves Instead of Degrades

A production applied AI system that doesn't improve over time doesn't stay neutral. Data distributions shift. Creator behaviors change. Market dynamics move. A system calibrated against last year's commerce patterns scores this year's creators against the wrong baseline.

The autonomous driving calibration loop makes the compounding effect concrete. A perception agent that misclassifies a scenario at 40 mph generates a calibration event. That event enters the training dataset. Every model trained after it handles that scenario differently. The system that drove 10 million miles is architecturally different from the one that drove 1 million. The models may be identical. The calibration data adjusted what those models learned to prioritize.

TikSense will calibrate against historical partnership outcome data: creators who were evaluated, partnered with, and produced measurable commerce results. That dataset accumulates value in a way no model upgrade replicates, because it encodes the actual relationship between pre-partnership signals and post-partnership outcomes. Without Calibration, the other six dimensions describe a snapshot. With it, they describe a system that compounds.

The Model Is Infrastructure. The System Around It Is the Product.

The Seven Dimensions of Applied AI is a practitioner framework. It describes what you design and control when building an applied AI product, not what you read about in model release announcements.

The dimensions don't operate in sequence. A Calibration update that shifts how the system weights a specific signal changes what Domain Logic treats as meaningful, which changes what Orchestration routes, which changes the context agents receive, which shifts where Governance & Containment flags disagreements, which changes what Human Judgment reviews, which generates the next round of Calibration data. Pull any one dimension and the others respond. The system is the interaction, not the sum of seven independent decisions.

This lives wherever you already are, Substack, X, or LinkedIn, drawn from building TikSense.