SHOWLABS
BRAND_DISCOVERY
A separate autonomous AI workflow tied to Shyft's retailer-bridge strategy. When users upload product data for unrecognized brands, the system triggers deep public-web research, identifies the brand from UPCs and SKUs, and feeds structured taxonomy data back into the product. The point is not enrichment. The point is growing the brand-retailer network automatically and creating a cross-sell story for unsigned brands.
OUTSIDE
THE USER LOOP
Most AI in products lives inside the user's chat window. This is different. The agent has a job, runs without being asked, and produces structured artifacts that improve the product as a side effect of people using it normally.
- Unknown uploads become growth signals
- Agentic research without user prompts
- Structured outputs feed product features directly
SYSTEM LAYERS
Autonomous Worker Pool
Parallel workers with lock-based coordination for safe concurrent research, so multiple discovery jobs run without duplicating each other's work.
Gemini Extraction
Gemini-driven extraction with strict validation, retry prompts, and canonical normalization — structured outputs the persistence layer can trust.
Taxonomy Repair
Mapping and repair loops over structured outputs, so categories stay consistent as the brand data scales past thousands of records.
Retailer-Bridge Integration
Discovered brands feed directly into the brand-retailer network model, improving UX on future uploads and creating cross-sell surfaces for unsigned brands.
DESIGN
PRINCIPLES
Autonomous pipelines fail differently than chat-style AI. These are the four design principles that made the difference between a research script and a system a product can rely on.
Agent Has a Job
The agent has a defined task rather than an open conversation. That changes everything about how it is designed, tested, and trusted.
Safety via Coordination
Lock-based coordination prevents duplicate work when many uploads hit at once, which is the practical risk of autonomous pipelines.
Structured by Contract
Every Gemini output is validated against a schema, retried on failure, and normalized before persistence — not whatever the model felt like returning.
Data as Product Surface
Every unknown upload becomes both a user-facing product event and a growth signal. The pipeline is not a data job, it is a product surface.