Back to projects

AdTech / Campaign Operations

AdOps Signal

Campaign operations product for AdTech teams managing delivery risk, policy ambiguity, brand-safety checks, and approval workflows.

Campaign operationsDelivery & pacingPolicy checksRisk scoringApproval workflowsAudit trail

Technical walkthrough available via GitHub profile or on request.

Strategic context

Why this matters

AdOps teams sit at the intersection of revenue, brand safety, and policy compliance, but they typically make high-stakes calls from scattered dashboards and tribal knowledge. AdOps Signal is a campaign-operations case study that shows how a senior product leader reduces decision latency without removing accountability: campaign signals are triaged automatically, policy is checked against a governed knowledge base, and every recommendation carries a confidence score and an audit trail. Nothing ships without a human approval on high-risk calls. This case study demonstrates product judgment on where AI should assist versus decide inside a media-operations workflow, and how to design for enterprise trust from the first version.

The problem

Product challenge

AdOps teams operate with scattered signals: campaign delivery gap, spend variance, publisher policy, brand safety, targeting setup, and creative or VAST tag issues. When these signals aren't connected, small problems compound into revenue loss, policy violations, client dissatisfaction, and manual escalation overload. Reviewing a single flagged campaign today means pulling data from a delivery dashboard, a policy wiki, a spend report, and a ticket queue, then applying judgment that varies by reviewer. The risk isn't just operational delay — it's inconsistent decisions on the exact area where AdOps organizations answer directly to legal, finance, and client trust teams.

Who this is for

Target users

AdOps Manager

Owns day-to-day campaign health and needs a single, trustworthy view of which campaigns require attention today.

Campaign Manager

Responsible for delivery and pacing on individual accounts, needs fast, explainable answers on why a campaign is flagged.

Product Manager for AdTech Platform (internal stakeholder)

Owns the roadmap for internal tooling and needs to justify where automation reduces cost without increasing risk.

Trust & Safety / Policy Reviewer (internal stakeholder)

Approves or rejects high-risk decisions and needs evidence, not just a verdict, to stay accountable to policy.

Revenue Operations Lead (internal stakeholder)

Cares about revenue at risk and SLA exposure across the full campaign portfolio, not a single flagged case.

Product bet

Product strategy

The product bet is that AdOps teams don't need more dashboards — they need a governed decision layer that turns scattered signals into one explainable, reviewable recommendation per flagged campaign, so decision quality improves without removing human accountability for revenue and brand-safety calls.

How it works

Product workflow

1

Campaign signal ingestion

Delivery pacing, spend variance, targeting configuration, and creative metadata are normalized into a single campaign signal record.

2

Policy / RAG check

Signals are checked against a retrieval-augmented index of publisher policy, brand-safety guidelines, and internal SLA documentation.

3

Risk classification

Deterministic rules and LLM reasoning combine to classify the issue by type and severity, with a confidence score attached.

4

Recommended action

The system proposes a specific next action — pause, adjust pacing, escalate to policy, or no action — with a plain-language rationale.

5

Human approval

Medium- and high-risk recommendations route to the right reviewer before any change is applied.

6

Audit log

Every signal, recommendation, reviewer decision, and outcome is recorded for governance and post-hoc review.

7

Impact tracking

Resolved cases feed back into a portfolio-level view of revenue protected, SLA exposure, and reviewer throughput.

AI leverage

Where AI is used — and where it isn't

AI is scoped in as one execution lever inside the product strategy, not the strategy itself. Every recommendation is explainable and every high-risk action is reviewable.

RAG over campaign and policy documents

Used only where explainability and traceability matter: publisher policy, brand-safety rules, and SLA terms are retrieved so recommendations cite the specific clause behind a decision.

Rules for deterministic checks

Hard thresholds (pacing bands, spend variance, SLA deadlines) are enforced with rules, not left to model judgment.

LLM for explanation and recommendation

The LLM synthesizes rule output and retrieved policy into a reviewer-readable explanation and a recommended next action.

Confidence scoring

Each recommendation carries a confidence score derived from signal completeness and policy-match strength, shown directly to the reviewer.

Human approval for high-risk decisions

Automation stops at the recommendation. Anything touching spend, brand safety, or client-facing delivery requires sign-off.

Audit trail for governance

Full decision provenance — inputs, retrieved policy, model output, reviewer action — is stored for compliance review.

Where AI is deliberately not used

AI is not used to execute pause, spend, or brand-safety actions autonomously, and it does not override hard policy or SLA thresholds — those stay deterministic rules. The model's role is confined to explanation, synthesis, and recommendation.

Product system

Core Product Modules

The product broken into its core operating modules — what each one does, who it's built for, and why it matters, before the architecture behind it.

Decision Layer

Campaign Risk Queue

Prioritizes campaigns with pacing, delivery, policy, or brand-safety risk.

User: AdOps Manager / Campaign Manager

Value: Reduces manual triage and highlights revenue-risk campaigns.

Review Layer

Policy & Brand-Safety Check

Checks campaign setup, creative signals, targeting, and publisher constraints against policy logic.

User: Policy Reviewer / AdOps Manager

Value: Makes approval decisions more consistent and explainable.

Operator View

Recommendation & Action Panel

Shows recommended next actions such as approve, hold, adjust targeting, review creative, or escalate.

User: Campaign Manager

Value: Converts risk signals into operational action.

Review Layer

Human Approval Workflow

Keeps high-risk decisions accountable through review, approval, and audit trail.

User: AdOps Lead / Reviewer

Value: Balances speed with brand, revenue, and compliance risk.

Insights Layer

Campaign Impact View

Links risk decisions to delivery gap, spend variance, SLA pressure, and affected campaigns.

User: Revenue Operations / Product Lead

Value: Connects AdOps decisions to business impact.

These modules define the product surface before the architecture and roadmap decisions.

How decisions flow

Product Architecture & Decision Pipeline

The architecture is built around one question a reviewer asks under time pressure: why is this campaign flagged, and what evidence backs the recommendation? Each stage exists to answer that question with less manual digging, not to automate the decision itself.

Stage 1

Inputs / Signals

Delivery pacing, spend variance, targeting, and policy/brand-safety signals

Stage 2

Processing / Decision Logic

Risk classification against policy and SLA thresholds

Stage 3

AI / Automation Leverage

RAG-grounded policy explanation with confidence scoring

Stage 4

Human Review / Governance

Reviewer approval required for medium- and high-risk actions

Stage 5

Product Action

Pause, adjust pacing, escalate, or approve the campaign

Stage 6

Measurement / Outcome

Revenue-at-risk reduction, SLA exposure, reviewer throughput

Under the hood

Technology Behind the Product

A compact view of what supports the product story — not a developer stack showcase.

Frontend / Product UI

Next.js, TypeScript

Keeps the reviewer-facing decision surface fast and type-safe as the workflow grows.

Backend / API

FastAPI

Serves signal ingestion and recommendation logic behind the review queue.

Data / Storage

PostgreSQL, pgvector

Stores structured campaign and audit data alongside the vector index used for policy retrieval.

AI / Retrieval

RAG, LLM abstraction

Grounds recommendations in retrieved policy text instead of unsourced model output.

Governance

Audit logs, approval workflows, risk scoring

Keeps every high-risk recommendation reviewable and attributable.

Deployment / Runtime

Vercel, Docker where applicable

Matches how the product UI and services would actually ship.

Product judgment

Key product decisions

The decisions below are the ones that actually shape whether a system like this earns trust in production.

1

Keep every medium- and high-risk recommendation behind human approval instead of full automation.

Why
An incorrect automated action on spend or brand safety is more costly and harder to reverse than a slower, reviewed one.
Tradeoff
Slower resolution than full automation, and it requires sustained reviewer capacity.
Impact
Builds the trust with operations and policy teams that adoption actually depends on.
2

Use RAG only where explainability and traceability matter, and keep deterministic rules for policy and SLA thresholds.

Why
Hard thresholds need to be enforced exactly; the model's value is explaining and synthesizing, not doing arithmetic on pacing bands.
Tradeoff
More implementation complexity than a single-model approach.
Impact
A more defensible system than a pure-LLM design, which matters when a reviewer has to justify a call to legal or finance.
3

Choose revenue-at-risk and SLA exposure reduced as the north-star metrics, not raw automation volume.

Why
AdOps and revenue leaders act on dollars and SLA exposure; a high 'cases auto-handled' number says nothing about whether the right calls were made.
Tradeoff
Requires explicit, labeled impact estimates rather than a simpler throughput count.
Impact
Keeps the roadmap pointed at the outcome the business actually cares about.
4

Design auditability into the schema from day one, not as a later add-on.

Why
AI-influenced decisions touching revenue and brand safety must be reviewable after the fact.
Tradeoff
More upfront data modeling effort before any feature work ships.
Impact
Removes a rework cycle later and signals enterprise-readiness to compliance stakeholders early.
5

Route decisions through the reviewer roles that already own the outcome, instead of introducing a new approval authority.

Why
Policy and revenue teams were unlikely to adopt a system that changed who was accountable for a call.
Tradeoff
Slower initial rollout while trust in the recommendations is established.
Impact
Lowers adoption risk by fitting into an existing accountability structure rather than fighting it.
6

Exclude autonomous execution of any campaign-changing action from MVP scope.

Why
Proving recommendation quality has to come before removing the human step, not after.
Tradeoff
The MVP moves slower than a fully automated pipeline would.
Impact
Keeps the first release low-risk enough to actually get approved for a real campaign book.

Scope discipline

MVP scope

What shipped in v1, and what was deliberately left out.

Included

  • Rule-based signal classification for pacing, spend, and SLA thresholds
  • RAG-backed policy explanation with citations
  • Manual reviewer approval queue for medium/high-risk cases
  • Portfolio-level revenue-at-risk and SLA dashboard

Excluded from v1

  • Autonomous execution of pause, adjust, or spend actions
  • Cross-campaign predictive risk modeling
  • Client-facing transparency reporting

Impact model

Measurement focus

What this system's output would be measured against once in operation — a measurement framework, not results from a live enterprise deployment.

Campaign review speed

How quickly a flagged campaign moves from signal to a reviewed, actionable recommendation.

High-risk decision coverage

Share of high-risk campaign signals that reach a reviewer with evidence attached, not just a raw alert.

Approval throughput

How many flagged campaigns a reviewer can move through once triage and evidence are pre-assembled.

Escalation reduction

Fewer tickets escalated without a clear next action already attached.

Risk signal distribution

Delivery pacing
Brand safety
Policy ambiguity
Spend variance

Roadmap

What comes next

V1

Governed triage MVP

  • Rule-based signal classification
  • RAG-backed policy explanation
  • Manual approval queue
V2

Portfolio-level decisioning

  • Cross-campaign risk correlation
  • Reviewer performance and SLA analytics
  • Configurable escalation policies per client
V3

Adaptive governance

  • Feedback loop from reviewer overrides into rule tuning
  • Predictive delivery-risk alerts before pacing breaks SLA
  • Client-facing transparency reporting

Product leadership takeaway

What this case study demonstrates

Product judgment

Recognizing that AdOps' real bottleneck was inconsistent judgment across reviewers, not a lack of data — and designing a system that standardizes the explanation, not just the alert.

Tradeoff accepted

Chose slower, human-approved decisions over faster full automation, accepting reviewer-capacity cost in exchange for defensibility on revenue- and brand-safety-critical calls.

Business relevance

Ties directly to revenue protection, SLA exposure, and the audit trail enterprise buyers and compliance teams require before they'll adopt any AI-assisted approval workflow.

Stakeholder complexity

Required aligning AdOps, campaign management, policy/trust & safety, and revenue operations stakeholders around one shared, explainable decision record instead of four separate views of the same campaign.

See how this pattern applies across other domains.

Every case study follows the same discipline: real problem, real users, explainable AI, and a measurable impact model.