03
Using Data and AI at Scale
AI scales as a system, not as a project: what one team builds, the next one reuses.
IV. Command Center · 03
03
AI scales as a system, not as a project: what one team builds, the next one reuses.
Scale is not doing more projects; it is making each project cost less than the one before. The first model pays for the platform, and the next ones are born with data ready, code standards, delivery pipelines and monitoring, without each team rebuilding the same foundation.
A platform alone does not scale. Scale comes when business, data, engineering and governance work as a single system: teams adopt the standard because it shortens delivery, the cost of every run is visible and every brand and channel uses the same base without losing what is its own.
Scale proves itself in the feature one team creates and another reuses, in the template the next model starts from, in the gateway that brings generative AI to every team with quotas and visible cost and in the customer recognised as one across six brands.

Lojas Renner
Grupo Boticário · GAVB · 2023
ChallengeCustomer data was scattered across brands and channels, with no single view or scalable segmentation to personalize communication within Brazil's data-protection law (LGPD).
SolutionA golden record by brand and corporate, a 360° customer view and AI personas in relationship cadences and remarketing; email campaigns brought in 30% more revenue across six brands and over 7 million customers.

Grupo Boticário
GAVB · 2023
ChallengeEach data science team rebuilt features in silos, with no consistency between training and inference, no governance and no reuse.
SolutionFeature groups by value stream, native and custom transformations, label generation, offline and online serving and drift detection, with scheduled pipelines.

GPA
Grupo Boticário · GAVB · 2021–2022
ChallengeRecommendations in the Cliente Mais and Clube Extra loyalty apps relied on a vendor due to be replaced by the third quarter, with little autonomy and off-strategy offers.
SolutionA four-wave pipeline of next best offer, basket-based cross-selling, in-category upselling and private labels, with final allocation by business rules; conversion reached 65% and 57% in the first releases.

Grupo Boticário
GAVB · 2021
ChallengeEach team built ML pipelines its own way, making it hard to productize, ensure data quality, monitor and document models.
SolutionAn internal library of pipeline components, covering tables, data quality and alerts, and a repository template with code structure, continuous delivery and business and technical documentation.

Grupo Boticário
GAVB · 2023
ChallengeTeams needed to experiment with LLMs under usage, cost and compliance governance, without scattered access keys.
SolutionA single generative AI API with team and global quotas, chat and embedding endpoints and ready-made examples for the internal hackathon.

Grupo Boticário
GAVB · 2023
ChallengeFranchise smart purchasing depended on forecasting sell-out by store and product; without it, purchase suggestions caused stockouts or overstocking.
SolutionA multi-brand forecasting pipeline by business unit, handling new stores, with a reliability layer, tracking indicators and data quality checks.

Grupo Boticário
GAVB · 2023
ChallengeMass communication and single-brand recommendation missed each consumer's interests across the four brands: they only got offers from brands they already bought.
SolutionCollaborative filtering on a consumer-by-brand-and-subcategory matrix, weighting store and e-commerce purchases, site and app interactions and CRM engagement, plus business rules.

Grupo Boticário
GAVB · 2021
ChallengeData teams needed shared standards for pipelines, SQL versioning, reference architectures and event contracts, without rework or inconsistency.
SolutionIngestion DataOps with versioned SQL, orchestration and continuous delivery by environment, reference architectures and data leadership ADRs, and an event schema registry.

Grupo Boticário
GAVB · 2022
ChallengeThe group's ML platform costs were growing with no visibility by pipeline.
SolutionPipelines that collect metadata from each ML run and a periodic cost report for the teams.