Inside the build
From market signal to product decision.
A closer look at the research, company knowledge, concept development, and review tools we built for FStea.
Start with evidence, not a blank prompt.
For FStea, deciding what to make next means connecting several different kinds of knowledge: what consumers discuss, what is already on shelves, what food-service brands are launching, and what the company can manufacture. Those inputs used to be easy to examine separately and hard to use together.
We built a bilingual product-development workspace that carries research through to a reviewable concept. The goal is practical: give the team a clearer basis for deciding which opportunities deserve development time.
The research exports contain 142,805 cleaned social-content records across Xiaohongshu, Douyin, and Weibo. That includes 114,292 distinct platform/content IDs; repeat captures are not counted as new unique posts. A separate analysis-ready export contains 99,520 records. Keeping those distinctions visible matters more than making every number look like a different result.
See the market from more than one angle.
The Trend Radar brings market signals into a reviewable list. The team can inspect the evidence, compare relevance, and approve, reject, rename, or rescore a trend. The archive preserves the analyst run so a promising idea does not disappear into yesterday’s chat.
Social attention is one input. The Competitive SKU Matrix brings in retail shelf records and marketplace data so the team can compare products, brands, and formats. Other collection modules cover tea-chain menus, search interest, seasonality, upstream ingredients, and regulatory developments. Availability depends on the source and collection setup; a configured connector is not a promise of continuous live coverage.
For example, an ingredient becoming popular online is more useful when the team can also examine where it appears on shelves, which formats carry it, and whether the timing fits an upcoming launch. The system keeps these inputs connected rather than flattening them into a single unexplained trend score.
Bring the company’s own knowledge into the decision.
A market opportunity only helps if the business can act on it. The product-asset matching layer connects opportunities to FStea’s capabilities, product information, and customer context. The Knowledge Vault makes routed company documents available to the tools that need them.
The Director is an AI advisor that can consult company knowledge, market evidence, and the current analysis run. Its answers include clickable citations when supported by sources, while inferred judgments are identified as judgments. The team can inspect where an answer came from instead of treating a fluent response as proof.
The Market Almanac organizes longer research into navigable chapters with citations and downloads. The Analysis Desk exposes the reasoning behind a run: rankings, challenges, research gaps, and the resulting memo. This makes the system useful for both a quick question and a deeper product discussion.
Move from an opportunity to a product dossier.
Territory Lab groups evidence into opportunity areas that people can review. A human decision sits between identifying an interesting territory and treating it as a direction to pursue. Reviewed opportunities can then move into the Realization Engine.
Directed concept runs and broader explorations build product dossiers with an evidence chain and a breakdown of fit. A four-persona advisor panel challenges each run from different perspectives, and the output identifies any advisor that did not contribute. The point is to expose trade-offs while an idea is still inexpensive to change.
Paid generation is controlled by explicit cost confirmation and spending limits. Runs progress through capped steps, save checkpoints, and support stopping after the current step. A stop does not pretend that work already in flight was free or undone. The team can review what exists and resume from a known point.
Develop the packaging—and keep the feedback.
The Design Suite takes a dossier that has passed the required gates and develops competing packaging directions. It keeps renders, critiques, feedback, and refinement hand-offs attached to the same product context. Packaging can evolve without losing the rationale behind the concept.
Customer, prototype, and launch feedback return through a shared feedback loop. The Experience Library keeps proposed and approved heuristics separate, and approvals record who made the decision. A Track Record view follows sealed predictions, deadlines, and final outcomes, so a confident forecast can be checked later.
These features support an iterative development process. They do not replace tasting, formulation work, feasibility checks, commercial judgment, or regulatory review. A generated concept is a proposal for the team to investigate—not a launch-ready product.
One workspace, with people in control.
The cockpit connects evidence, territories, concepts, design, and feedback through a shared navigation path. English and Chinese interfaces, a searchable glossary, and a command palette help different members of the team work from the same vocabulary.
Sensitive tools and company records are restricted on hosted deployments. Human approvals remain part of the workflow, and source status and incomplete results are surfaced rather than silently presented as complete. The result is a system built around a real decision process, with room for the team’s expertise at each step.
The small generator below illustrates one part of that experience: choosing a category, channel, format, and direction, then examining a concept and its reasoning. It uses fixed sample data and illustrative scores. The full FStea workspace described here has a much wider research and review workflow.