Ugo ChukwuAI & Financial Systems
← Work

Applied AI / Strategy & Creative

Strategy & Copy Agent

An evidence-bound workspace that turns research into distinct campaign strategies, hooks, copy and short-video scripts — with immutable revisions, exact-version approval and a handoff that cannot run before a reviewer signs it.

StatusValidated prototype
DomainCampaign strategy & copy
RoleArchitecture & build
Core stackPython · FastAPI · React

An independent sanitized reconstruction, not Peerless production. All fixtures, brand identities, observations and sources are synthetic.

01 — The problem

Ask a model for five strategies and you get one strategy five times.

Generative copy tools collapse toward the mean. Prompted for options they produce variations in wording rather than genuinely distinct positions, which leaves a strategist choosing between synonyms and calling it a portfolio.

The second failure is provenance. Copy that cannot point back to the research finding that justifies it is unreviewable — and unsafe, because nothing stops a confident claim about a product from being invented at the sentence level.

Cost of the gap

A portfolio of options that are all the same bet.

Risk

Product claims that no evidence supports reaching a live campaign.

Blocker

No record of which exact version a reviewer actually approved.

02 — The system

Distinct by check, evidence-linked by contract, approved by version.

Strategies are generated as a portfolio and then checked deterministically for coverage and diversity, so genuinely distinct positions survive and near-duplicates do not. Every strategy contract and copy asset links to the evidence behind it. Revisions are immutable, approval targets one exact version, and handoff is idempotent and gated on that approval.

Portfolio coverage checks

Deterministic diversity checks reject a portfolio whose options are variations on one idea.

Evidence-linked contracts

Each strategy and copy asset carries the research claims that justify it.

Safe-claim checks

Copy is checked against the approved claim set before it can be handed off.

Immutable revisions

Edits create new versions. History is never rewritten.

Distinct roles

Strategy lead and brand reviewer hold different permissions, enforced server-side.

Idempotent handoff

Handoff runs once per approved version and rejects anything pre-approval.

03 — Architecture

How a strategy reaches handoff

Approved researchBrand rulesSafe claim set
Portfolio generationdistinct positions · coverage check
Evidence bindingstrategy contract · claim provenance
Copy & script assetshooks · copy · short-video scripts
Review & approvalimmutable revisions · exact-version approval
Approved strategyCopy assetsIdempotent handoff
Strategy and copy agent showing a portfolio of distinct campaign strategies
Strategy portfolio · synthetic brand

04 — Engineering decisions

Where the judgment actually lives.

Diversity is checked, not requested

Asking a model for "five different angles" produces five similar angles. A deterministic coverage check over the generated portfolio is the only reliable way to know the options are actually distinct.

Claims are a bounded set

Copy may only assert what the approved claim set permits. This is the difference between a creative tool and a regulatory incident.

Immutable revisions, exact-version approval

A reviewer approves a specific version, not a document that keeps changing underneath them. Approval that can drift is not approval.

Optimistic concurrency on every mutation

Stale writes are rejected rather than merged. Two people editing a strategy is normal; silently losing one of their edits is not.

Roles are server-derived

Tenant scope comes from a signed session, not from a client-supplied identifier. Permission checks that trust the caller are decoration.

Useful without a provider

The deterministic path produces a complete, reviewable workflow with no model key — which is what makes the review behaviour independently testable.

05 — Reliability & controls

What the system does when things go wrong.

Failure modeBehaviour
Stale version submittedRejected by optimistic concurrency; the editor is shown the current version rather than overwriting it.
Handoff before approvalRejected outright. Covered by a dedicated pre-approval handoff test.
Duplicate handoffIdempotent — the second call returns the first result rather than repeating the work.
Tampered session tokenDenied; covered by token-tampering tests.
Wrong role attempts approvalDenied. Only a brand reviewer can approve, and only one exact pending version.
Service restart mid-workflowState is durable; covered by restart-durability tests.

06 — Evidence

Project statusValidated prototype

Verified results

None published. No client campaign results are claimed on this page.

Benchmarked results

Public dataset: Deterministic evaluation set, local and CI verified

Delivery tier: Tier 1 runnable repository

Evaluated: portfolio diversity · provenance · belief movement

Also evaluated: success-signal completeness · safe copy · synthetic labelling

The committed test suite covers UI interactions, tenant isolation, token tampering, role denial, stale writes, restart durability, exact-version approval, pre-approval handoff rejection and idempotency. These assert behaviour, not creative quality.

Simulated results

All public fixtures, brand identities, observations, sources, screenshots and evaluation outputs are synthetic.

Projected business value

Not claimed. Historical production measurements and private implementation details are deliberately excluded from this project.

Verified, benchmarked, simulated and projected figures are kept separate on purpose. A benchmarked number is never presented as a verified client result.

07 — Technology

PythonFastAPIReactTypeScriptSQLiteOpenAPIpytest

Technology supports the story. The architecture and the controls are the story.

Generating copy nobody can approve with confidence?

We can bind every asset to its evidence, bound the claims it may make, and make approval mean one exact version.

Book a Discovery Call