01
A manager named the pain.
He was describing a project workflow with scattered content requests, market nuance, and review complexity. I said, "Wait, I think I can build that."
Product design + strategy / Exploratory POC
Content Operations Agent
I turned a manager pain point into a working AI workflow that earned executive validation and funding.
Role
Initiated, designed, built
Timeline
3-week MVP
Stack
React, TypeScript, Node
Outcome
Funded global rollout
Demo moment showing the POC's agent workflow in motion.
Agent workflow demoThe Story
This did not start as a formal brief. It started as a conversation where I heard a real workflow gap and saw a way to prototype something useful.
02
I learned from the people closest to the work.
I talked with UX designers who deeply understood the users and the existing internal website, then used those conversations to understand what the POC needed to prove.
03
I shaped the POC with real user context.
I listened for where authors, reviewers, and content owners needed confidence: intake clarity, market-specific guidance, consistent language, compliance visibility, and a human decision point.
04
The prototype created a path forward.
The proof-of-concept gave leadership something concrete to evaluate: a working product direction, not a slide deck or abstract AI strategy.
The Problem I Heard
Workflow
Content work was too distributed to manage cleanly.
Content could involve HR partners, comms leads, market teams, source documents, templates, and reviewers across regions.
User risk
Small inconsistencies created trust issues.
- Wrong locale formatting
- Inconsistent spelling or tone
- Policy language drifting by market
Before vs. After
Legacy framework
2-3 weeks per article
- 01 / Manual briefs
- 02 / Decentralized emails
- 03 / Siloed market translation
- 04 / Manual policy compliance review
Orchestrated framework
Under 15 minutes
- 01 / Unified intake
- 02 / Parallel multi-agent validation
- 03 / Single human-in-the-loop review pipeline
Why it matters
Framing the speedup next to the problem grounds the engineering and interaction decisions in business value: less coordination drag, faster market localization, and one clearer path from brief to review-ready article.
The POC I Built
Intake
Capture the brief, sector, market, audience, and source context in one place.
Agent workflow
Draft, check, and revise content against market and compliance needs.
Human review
Keep the final decision with the reviewer, with flags and rationale visible.
The Multi-Agent Architecture
I architected the interaction model around a visible orchestrator, giving human authors transparent control over modular AI sub-processes instead of hiding automation inside a black box.
Core Orchestrator
Coordinates the agent sequence, preserves handoffs, and keeps the author aware of what is happening at each step.
Intake
Checks that the brief has the minimum context needed to begin.
Clarifier
Surfaces missing details before the system commits to generation.
Router
Directs the request to the right sector, market, audience, and source logic.
Market
Adapts content against local conventions, language, and operating context.
Compliance
Flags policy-sensitive content so reviewers can see risk before publication.
Revision
Turns critique into an updated draft while preserving reviewer intent.
Translation
Prepares localized variants once the source content is approved.
Demo clip showing the POC moving from brief to review-ready content.
POC review demo / 2xKey Product Decisions
01 / Permission model
Separate admin governance from author workflows
Problem
A single view mixed enterprise governance with everyday article management.
Move
Split the experience by permission level: admins get system-wide controls; writers and managers get focused ownership tools.
Result
Each role sees the right scope without extra operational noise.
Admin view
Monitor article health, spot stale or low-performing content, and alert owners to review, update, consolidate, or remove.
Non-admin view
Manage owned and direct-report articles, track stale reviews, filter by owner/status, transfer ownership, and make updates.
02 / IA model
Sector before market, not market alone
Problem
Design conversations made it clear that market-only organization ignored how content ownership actually works.
Move
Made sector the primary organizing tier across profiles, filters, and articles.
Result
Authors choose "PFNA → United States," so the POC feels specific to PepsiCo's operating model.
Technical Alignment
Ensured data indexing paths match PepsiCo's operational hierarchy, mapping API calls directly to sector boundaries.
03 / Workflow surface
Merge review queue into content library
Problem
Users needed to understand what was in progress, what needed attention, and what was already live.
Move
Collapsed review and library into one All Articles surface with three tabs.
Result
Reviewers and authors can move between work-in-progress and live content without losing context.
Technical Alignment
Optimized state management in React to prevent rendering lag when toggling between dense, real-time localized dataset tables.
04 / Grounding model
Sources as profile data, not per-article attachments
Problem
Repeated source setup would make the POC feel slower than the workflow it was meant to improve.
Move
Put sources on sector, market, and audience profiles instead.
Result
Grounding follows ownership: PFNA sources guide PFNA content; US sources guide US procedures.
Technical Alignment
Caching system data globally on sector profiles drastically reduced payload sizes and context-window token overhead for subsequent Claude API inference cycles.
Outcome + What's Next
Current outcome
The functional proof-of-concept passed executive validation and secured corporate funding for a centralized global rollout.
Next validation
Architecting the phase-one production pipeline, establishing LLM evaluation metrics for market-specific compliance accuracy, and kickstarting engineering integration.