Working systems, grouped by operating problem and proof depth.
Showing 22 systems
Compare proof depth, problem space, source status, and the best entry point.
Planners get transfer recommendations before stockouts hit, with a 50-warehouse, two-thousand-SKU, 100-scenario run proven under 18 seconds.
Turns failed deliveries into owned claims with evidence, exports, deadlines, and audit history. Avoids roughly €32K in manual case chasing.
The entity snapshot and hash-chain essays show how legal documents keep identity, numbering, compliance sidecars, and audit proof separate by entity.
Replaced three carrier portals with one API and live tracking stream. Built for 100 active shipments, 500 events/minute, and sub-200ms fanout.
Turns scattered finance exceptions into one owned queue with workflow resolution and audit export. Targets $40K+ annual recovery for SaaS operators.
Connects missed vendor obligations to a collectible settlement process with paired ledger discipline.
Turns procurement signals into frozen alert history, risk scoring, and reviewable supplier evidence.
Shows how AI extraction stays behind human confirmation before it becomes business truth.
Shows how five resolution strategies move invoice matching from manual review to controlled exception handling.
Shows the ingestion side of event design under memory, throughput, and failure-isolation constraints.
Grounds event infrastructure in a revenue workflow where payment states and retries affect churn.
Clinic operators recover lost appointment capacity with live slot computation, cancellation backfill, and database-enforced double-booking protection.
Defends append-only records, sequence ordering, and the constraints needed when history must be defensible.
The delivery security essay proves the system was designed around hostile URLs, retries, and destination safety.
Shows how dynamic DAG branching was handled without turning workflow definitions into path explosion.
Best proof for optional Kafka, direct HTTP intake, rate limits, templates, and tenant boundaries.
The optional knowledge graph essay shows where model ambition met memory cost and retrieval value.
Best proof for messy settlement dates, confidence scoring, and why exact matching breaks real finance workflows.
Searchable operational knowledge layer
Mar 2026 · 6 min case study
Shows how replicated property data became a queryable intelligence layer without calling four services per question.
Manual Idealo price checks covered 40 products per session at 45 minutes each. Automated ranking-aware pricing raised rank-1 coverage from 45% to 70%.
Useful proof for modeling around survivor bias and missing data instead of pretending the dataset is clean.
A 4,162-product Shopify catalog ran on 90 minutes of daily manual price updates across 13 suppliers. A 7-minute pipeline replaced it end to end.
Get Notified
New system breakdown? You'll know first.