Executive summary
IPG needed to scale the research behind product listing improvements. Awaken built its AI commerce system and delivered a working MVP in 90 days. The platform processes products in parallel and produces branded optimization briefs for strategist review.
The situation
IPG was a Fortune 400 group serving 3,000 brands, including Kenvue and General Mills. The project needed to support commerce work across thousands of products and multiple retailers.
IPG reported $10.7 billion in total revenue for 2024, including billable expenses.1
The challenge
Manual listing research took days and varied by analyst. Recommendations had to balance search visibility, persuasive content, product accuracy, and brand restrictions.
A useful search keyword might imply a claim the brand cannot make, and a persuasive description might fail a retailer’s formatting rules. The system needed to account for both when turning research into recommendations.
What Awaken built
From a product identifier, the system combines product facts, competitor research, internal Amazon keywords, and competitor keywords. It generates titles, bullets, descriptions, specifications, and image guidance, applying approved claims, legal restrictions, product-line limits, and retailer requirements.
Coordinating the work
Roughly fourteen LangGraph ReAct nodes structure inputs, generate recommendations, validate them, and score outputs. Agents can call tools during execution.
The graph passes context between steps and separates generation from evaluation so drafts can be revised in response to feedback. LangGraph supports this combination of fixed steps and agent decisions.2
Checking content and data
Deterministic checks cover length, prohibited terms, keyword coverage, and required fields. Model review supplies qualitative feedback, and failed validations trigger revisions. Pydantic schemas and reconstruction handle malformed outputs. Engineers can inspect execution through LangSmith traces, while HHEM-2.1 screens for factual consistency.
We kept format and factual checks separate because a response can pass one and fail the other. That helps the workflow and its engineers identify what needs fixing.
HHEM evaluates whether a generated statement is supported by the supplied source. Its score helps identify content that needs attention during review.3
How we worked
IPG engaged Awaken directly. Hanzhi Guo built the system and led the team from zero to MVP in 90 days. Awaken owned the implementation and delivery; IPG’s strategists retained responsibility for reviewing the resulting briefs and deciding which recommendations to use.
Results
Awaken delivered the working MVP in 90 days. It processes products concurrently and produces consistent briefs with traceable recommendations. IPG can configure retailer rules and choose models for individual tasks, then reuse the workflow across its product work.
What changed for the business
IPG’s strategists can start with a prepared brief, review the proposed changes, and decide what to publish. The delivery team can trace how each recommendation was produced when maintaining the system.
Sources and methodology
- IPG, fourth-quarter and full-year 2024 results. Historical company revenue, including billable expenses.
- LangGraph, workflows and agents. Background on orchestration and evaluation patterns.
- Vectara, HHEM model documentation. Background on factual-consistency screening.
What could we build for you?
Tell us about your project. Let’s talk through what it would take.
Let’s talk