Behavioral Inventory Demand Forecasting

How the system actually runs.

Each stage below is a real component in the deployed pipeline. Stages outlined in teal are control points: the places where the architecture constrains what the model can receive or emit.

What it runs on.

How it is wired.

1. Fully cloud-native on AWS with no on-premises footprint.
2. A daily scraping pipeline ingests TikTok trend data, passes the content through Bedrock Claude for scoring, and writes the normalized signals to the S3 feature store.
3. A Redshift pipeline merges historical sales, the promotional calendar, and supplier lead times into the same feature set.
4. The SageMaker endpoint consumes the merged feature set and runs daily reforecasts across all SKUs. The endpoint is VPC-internal with no public ingress.
5. All service-to-service calls use TLS 1.3. No data exits the private subnet at any point in the pipeline.

What keeps it inside its contract.

Guardrails were designed in from the start, not added after the fact. Where the system touches untrusted input, a control sits between that input and the model.

• Red-team exercises explicitly tested prompt injection via scraped product description text embedded in social content.
• A sanitization layer strips instruction-like patterns, including imperative verb structures, role-assignment strings, and “ignore previous” variants, before any text reaches the Bedrock endpoint.
• The system prompt is locked at the infrastructure level. No user-supplied input path exists in this pipeline.
• Output validation enforces that every Claude response is a float between 0 and 1 before it enters the feature store.
• Out-of-contract outputs, meaning non-float, out-of-range, or null, fire an alert to the ML ops team, and the prior day score carries forward automatically.
• No PII transits the Bedrock endpoint. Only product category labels and aggregated trend text are passed.

What we still track.

No production AI system is finished. These are the live risks carried by this architecture, documented for the operating team rather than buried.

• Adversarial content embedded in scraped social posts could attempt to manipulate trend scores if sanitization gaps exist.
• Model drift on the Chronos endpoint if the product catalog expands rapidly without scheduled retraining.
• Supplier lead-time data freshness. Stale ERP records degrade forecast quality without a clear signal to operators.
• Over-reliance on trend scores for niche SKUs where social signal volume is too thin to be statistically meaningful.

What changed.

• Stockout rate on trending SKUs reduced measurably in the first full quarter post-deployment.
• Overstock markdown losses declined through better category-level suppression of slow movers before they aged into markdowns.
• The ML ops team gained a clear alerting path for when the AI component operates outside its expected output contract.
• The architecture established a reusable internal pattern for safely passing scraped unstructured data through an LLM enrichment step.

On numbers. Client identity and any identifying detail are withheld under NDA. Outcomes are stated as direction and materiality rather than as figures we are not free to publish.