In Mexico, nobody was tracking its prices.
Now 160 retailers are checked up to six times a day.

Retailer formats vary, but the output has to match the company's existing pricing schema exactly, or nothing downstream works:

• Scheduled jobs trigger Lambda orchestration and containerized processing on ECS/Fargate, deployed across U.S. East, Canada Central and Mexico coverage regions in private subnets.
• Amazon Bedrock extracts the product URL, listed price, availability, ratings, buy-box winner and first- or third-party seller classification.
• The pipeline captures screenshots as evidence and normalizes the output to the existing pricing schema.
• EFS holds temporary processing files, and the architecture connects processing to S3 outputs, a metadata store and the analyst application.

Why: a model that reads pages copes with layouts changing, and a schema that never changes is what keeps the reports working.

The costly mistake in pricing intelligence is comparing the wrong product. So a human review workflow lets the team correct and approve AI-suggested SKU matches before they're used. Consolidated records then flow into Snowflake, keeping schema compatibility while adding seller-location and marketplace information the old tool never had.

Coverage grows when the business can add retailers without engineers. A configuration interface lets team members add a retailer and test extraction with a dry run before it joins the scheduled service, with the required scraping parameters identified automatically.

The service grew to 25 Mexican retailers with 200 SKUs, 100 U.S. retailers with 600 SKUs and 35 Canadian retailers with 350 SKUs. Large retailers are refreshed six times a day and smaller ones three times.

Pricing was one part of a wider program. The company's device estate generates about 70,000 telemetry records a minute, but product teams couldn't investigate mission success, coverage and performance without writing SQL. An IoT insights proof of concept connected data preparation to questions and reports:

• Raw S3 data passes through Glue ETL into curated datasets, with the Glue Data Catalog and Athena for analytical queries.
• Amazon Q Business uses the S3 knowledge source, with a device-specific vocabulary and role-based access for each product team.
• QuickSight dashboards with Q show mission success, coverage and performance trends, through charts and natural-language questions.
• Python, Pandas and Plotly produce daily and weekly summaries, with PDF and Excel exports for wider distribution.

Devices running outside normal parameters went unnoticed until a customer complained. A separate anomaly-detection proof of concept tested how the same device data could flag failure patterns earlier:

• SageMaker runs time-series models and failure-pattern clustering. Its Model Registry records versions, and endpoints serve inference.
• The architecture traces device data through IoT Core and S3 to the inference engine.
• Classified results feed QuickSight, and SNS routes notifications by issue type and severity.
• Endpoint monitoring and baseline jobs send operational metrics to CloudWatch, including model drift, so alert quality doesn't fall as the fleet grows.
• Maintenance predictions from usage patterns were explored too.

These are proof-of-concept (POC) capabilities, and the operational impact across the fleet is still to be measured.

The consumer insights team commissioned every study from outside agencies, which limited how often it could test an idea. The synthetic-persona platform uses the company's existing consumer research and survey data to derive demographic distributions, ownership patterns and behavior, with study-screening criteria mapped to the generated personas:

• Each persona and the complete questionnaire go to the Bedrock inference layer together, so the full study context is available while it answers, and each persona answers consistently across the survey.
• Parallel execution supports larger simulations.
• Aggregated responses are organized into six research dimensions: role in category, white space, psychological drivers, friction and unmet needs, premium preferences, and demographic sensitivity.
• A React interface lets analysts adjust screener mappings, search earlier studies and validate outputs, in English, French and German.

The output is exploratory hypotheses from synthetic personas, used to decide what to study with real consumers, in days rather than weeks. It doesn't replace the real studies. Synthetic research is useful as a fast first pass, and only when it's labeled as one.

Customers find recipes in social video and then have to adapt them to their own appliances. The recipe proof of value tackled that:

• Transcribe extracts the spoken content from a video, and Bedrock structures the recipe and adapts it to the selected appliance and dietary preferences, including gluten-free, dairy-free, vegan, keto and paleo.
• Appliance manuals supply what each appliance can do, and OpenSearch holds searchable recipe, ingredient and appliance representations.
• A Streamlit interface accepts a video or URL, the appliance and the preferences, and a chat interface handles further changes, with FastAPI on EC2 for iterative testing.
• The prototype includes ingredient substitutions, allergen detection and nutritional information.