System Map

From disparate data to better product decisions.

The data behind food and beverage innovation sits in sources that were never built to talk to each other. We collect consumer, retail, product and enterprise data, process each on its own terms, and connect the results into one body of evidence our category experts and AI can reason over together. Every decision made on it returns as a signal that sharpens the next one.
01 — Ingestion & Processing
Three pipelines, one shape
Each source follows the same three-part pattern — collect the raw material, resolve and enrich it into structured records we can trust, then analyze and model it into something with predictive value. What differs is the work inside each step, because a video, a sales record and a product label are not alike.
CollectGet the raw material in
Resolve & enrichTurn it into trustworthy structure
Analyze & modelMake it predictive
Consumer
  • Social video & conversation
  • Creator & audience signals
  • Survey panel
45M+ unique posts · continuous daily ingestion
Gather & transcribe
Food and beverage social content is gathered continuously. Audio is separated from video and transcribed, so what people say becomes searchable text alongside what they caption — then quality-filtered to drop unusable transcripts.
Resolve language to concepts
Captions and transcripts are broken into terms and converted into numeric representations of meaning, then matched against the ontology by similarity. A post gets tagged with the concepts and attributes it genuinely discusses, not the exact words it happened to use.
Score, cluster & read
Each concept earns a daily score weighted by real engagement, not raw views. Posts are then clustered into themes and read by AI to pull out what consumers are actually saying, in their own words. Survey waves are modeled two ways — each wave alone, and rolling across waves — to predict demand by demographic segment.
Retail
  • Syndicated sales & velocity
  • Distribution & channel
  • Pricing & promotions
  • Consumer receipts
Weekly sales history · basket-level purchase detail
Ingest sales & baskets
Two very different records of what actually sold: syndicated point-of-sale history covering revenue, velocity and distribution by channel, and real consumer receipts broken down to the individual line item.
Resolve to the product spine
A sales row and a receipt line are useless until you know exactly which product they describe. Each is matched to our catalog in three passes — product code, then brand and name, then brand, name and price together, which is what inherits the product's full attribute profile. Every run reports its own match quality.
Forecast concept success
Matched sales history becomes a modeling panel in which attributes are the features — so the models learn which attributes actually move velocity, and can forecast how an unlaunched concept would perform.
Product
  • Ingredients & nutrition
  • Claims & certifications
  • Formats, sizes & packaging
  • Reviews & ratings
576K products · 10 grocery retailers
Collect the shelf
Product data is collected across ten grocery retailers on a quarterly cadence and normalized into a common set of product tables, with changes tracked between cycles so price and assortment history accumulate rather than overwrite.
Extract & verify attributes
The hard part, and the core of the platform. Attributes are pulled from three places at once — retailer-declared fields, free-text descriptions, and AI reading the product label images — then checked against a human-vetted register of known attributes. Only register-matched values are marked verified and used downstream; anything unrecognized is held in a review queue for a person to rule on, rather than discarded or silently promoted.
Mine reviews per product
Reviews attached to those products run through sentiment, theme discovery, usefulness scoring and language-intensity analysis — turning thousands of opinions into what consumers specifically praise and complain about, product by product and attribute by attribute.
Enterprise
  • Past launches
  • Innovation decisions
  • Commercial outcomes
  • Manufacturing constraints
Joins directly at the connected layer
Enterprise decision data needs no collection or enrichment pipeline of its own. A past launch is already a product with attributes and a commercial outcome, so it attaches straight onto the product spine and inherits the same attribute vocabulary as everything else. What it adds is the one thing no external dataset contains: the decision that was made, and what happened next.
Our own brands first, then each partner engagement
Everything converges on one object
02 — The Connected Layer
The attribute is the unit of analysis

A concept in the ontology maps to an attribute on a real product — the connection no single dataset can make on its own.

The attribute is the platform's unit of analysis — the one object that appears in every source we touch. The two engines below exist to serve it, working inward from opposite directions: one turns what consumers say into attributes, the other attaches what actually sold to them.
Consumer side
The language engine
Resolves messy consumer language into attributes.
Our proprietary food and beverage ontology: every concept, every real-world way people phrase it, and how those concepts relate. "No dairy", "dairy free" and "non-dairy" resolve to one idea rather than being counted as three separate trends. Curated over years, with AI now proposing new concepts and relationships for expert verification.
87.1K
Real-world phrasings
52.7K
Synonym sets
The shared object
The attribute
organichigh proteinno dairysingle serve
An attribute is small enough to be read off a product label, spoken in a video, declared by a retailer and carried on a sales record — which makes it the only object that exists in every source we touch. It is what we resolve language into, what we attach performance to, and what a product concept is ultimately composed of.
9.8K
Canonical attributes in the expert curated master
24
Attribute categories
Commercial side
The identity spine
Hangs commercial evidence on the right attribute.
A single catalog of what actually exists on shelf, and the machinery to match anything to it. Sales rows, receipt lines, reviews and past launches all resolve to the same product record — by product code, then name, then attribute and price similarity — so every attribute steadily accumulates real evidence of how it performs.
576K
Cataloged products
10
Grocery retailers
Why our inputs to AI are different
A general-purpose model can reason, but it cannot do any of this for itself. It cannot assemble years of proprietary consumer, retail and product data, or resolve ten retailers' inconsistent language into one vetted vocabulary. It cannot know how attributes relate to one another, which ones a category expert would actually trust, or which ones measurably move velocity — because that knowledge does not exist on the open internet. It lives in what we have built. The model reasons over that work; it is no substitute for it.
The connected layer produces
03 — Insights & Functionality
What the connected data becomes
Trends
  • Where a concept sits across an 8-stage lifecycle
  • Durable trend vs. short-lived spike — with the reason
  • Macro-trends, not isolated keywords
Consumer Analysis
  • Themes and sentiment across social and reviews
  • Consumers' own words, quoted and ranked
  • Demand prediction by demographic segment
Retail Analysis
  • Distribution, velocity and category benchmarks
  • Pricing down to the individual store
  • Where a category has attribute white space
Product Predictions
  • Pre-launch velocity, revenue and distribution
  • Which attributes actually drive performance
  • Ranked concepts with confidence ranges
Business Context
  • Brand goals and guardrails
  • Manufacturing and cost constraints
  • Prior decisions and what they returned
Insights become the context for reasoning
04 — Reasoning & Decisions
Where experts and AI meet the data
User Inputs
What our experts put in
Research questions
Product concepts
Business constraints
Business goals
Starday's category experts translate a client engagement into the questions, concepts and constraints the platform works from.
AI Reasoning
Reasoning over connected context
The question meets our insights, not the open internet. The system retrieves the relevant evidence from the connected layer, applies the questions, tradeoffs and constraints our category experts have encoded, and surfaces patterns across sources that no single dataset shows on its own.
Grounded in our data
Expert judgment encoded
Traceable to source
Outputs
Decision recommendations
Key insights
Recommendations
Product concepts
Performance forecasts
Which opportunity to pursue, what to build, and how to build it.
Decisions return as signals
  • Which insights proved useful
  • Which concepts were prioritized
  • Which attributes were chosen
  • Which products got built
  • What happened in market
  • Forecast vs. actual outcome
Proprietary decision data no one else has.
Signals return into the enterprise data and retrain the models — every engagement teaches the platform how better product decisions are made.
Starday Intelligence Platform — how it works
Figures are current platform counts