WORK / CONSTRUCTION / BUILDING PRODUCTS
DMC Workflow Automation
BLOX Built: Stood up automated workflows across the Design-Manufacture-Construct (DMC) stack — anchored by a model-routed data clerk that completes canonical part records for tens of thousands of SKUs using the enterprise evidence already in the system. A tiered cascade (gpt-5.4-mini → gpt-5.4) classifies, enriches, and validates each part at roughly seven cents a pop, with human approval and audit trail as first-class surfaces.
Metric: ~$0.07 per part · end-to-end enrichment. Timeframe: Discovery through live rollout.
Shipped a tiered-cascade Python pipeline routing each part through the cheapest model that survives the quality gate. Pass 1 classifies to CSI MasterFormat with gpt-5.4-mini and structured output. Pass 2 enriches against Exa + Firecrawl grounded web context, biased toward linked datasheets; low-confidence records escalate to gpt-5.4. Pass 3 runs a multimodal vision check on gpt-5.4 to validate product images against the enriched description. Every pass emits a confidence score and a review note. Records flow into a human-approval queue with immutable audit trail. The orchestrator enforces a batch budget ceiling and degrades gracefully when any signal is weak.
This public case-study shell gives crawlers the core company, industry, outcome, and implementation context before the React app renders the full interactive detail page.