Evidence first
An advertising claim cannot enter creative output without attributable supporting evidence.
03 Agentic media workflow · 2026
Affinery.
An evidence-backed, human-approved workflow for researching affiliate opportunities and preparing professional TikTok-ready product media.
01
Operator console
Running-state capture
Research, review, publishing and learning in one operator workspaceMock data · 12 Aug 2026
02
Product walkthrough
16 running-state viewsOne controlled loop.
Every handoff visible.
The interface exposes what the workflow is doing, what evidence supports it and where an operator must decide. These captures use an isolated mock database, so no live records were changed.
01 Command and campaigns
Intake · budgets · tracesThe command center summarizes recorded performance and pipeline health. Campaign views make budgets, costs, creative counts and every agent artifact inspectable.
02 Product intelligence
Inventory · evidence · claimsProducts stay connected to attributable sources, approved claims and asset rights. Research can be audited before any statement reaches a script.
03 Human review
Render · disclosures · decisionGenerated media cannot publish itself. The review queue brings the render, verified claims, cleared rights and required disclosures together for an explicit approve, revise or discard decision.
04 Publishing and learning
Drafts · metrics · commissionApproved content enters an auditable publishing queue. Recorded TikTok metrics and affiliate conversions feed the learning loop without inventing engagement when no snapshot exists.
05 Native operations
Jobs · agents · Flow handoffAffinery owns its durable jobs, schedules, retries and telemetry. Bounded agents expose their process, while the guarded Google Flow handoff returns a render candidate to the same review boundary.
03
Project overview
Auditable automationAutomate the workflow, not accountability.
Affinery turns a constrained niche into a traceable production loop: product research, evidence verification, opportunity ranking, creative strategy, scripts, assets, media generation, quality checks, human review and publishing.
The runnable build includes database-backed campaigns, products, evidence, claims, review decisions, publishing records, metrics, automation jobs and agent telemetry. External providers remain explicit integrations rather than mocked production claims.
Human authority
Agents may research and prepare work, but only an operator can approve public publishing.
Bounded agents
Budgets, typed contracts, retries, idempotency and audit records constrain every automated stage.
04
Technology stack
Runnable foundationDesigned as
a controlled loop.
The current slice runs locally with explicit boundaries for retrieval, media generation and TikTok publishing. Production identity, PostgreSQL/pgvector and provider credentials remain deployment work—not portfolio claims.
Python 3.12
Typed workflow contracts, orchestration, QA, review and supervision logic.
FastAPI + Pydantic
Operator routes, validated actions and the browser-facing application boundary.
SQLite
The current source of truth for campaigns, evidence, versions, jobs, approvals and metrics.
Bounded agents
Codex-led orchestration and narrow worker contracts with budgets, retries and telemetry.
Google Flow handoff
A guarded operator-assisted browser lane packages requests and returns checked MP4 candidates.
TikTok adapter
An approval-gated publisher queue with explicit OAuth and verified-media requirements for production.
05
Current result
Runnable control planeThe repository delivers the complete database-backed state flow with deterministic fixtures, while clearly separating working local behavior from credentials and infrastructure still required for production.
16Captured operator views
01Human-only approval boundary
60/30/10Weekly learning allocation
Next case / 04