Blog
AI agents, RAG, and integrations — notes from production.
General
Resilient Third-Party API Integration
Third-party API integration that survives outages: one client per vendor, timeouts, safe retries with backoff, a queue for slow work, and reconciliation.
General
Multi-Tenant RAG for B2B Products
Multi-tenant RAG in a B2B product: filter retrieval by tenant in the database, mirror document permissions, and keep caches, logs, and deletions scoped.
General
Usage-Based Billing with Orb
Usage-based billing with Orb: send events through an outbox, use stable idempotency keys, respect the grace period, and correct mistakes without losing data.
General
Chargebee Plan Changes and Proration
Chargebee plan changes done right: update_for_items, proration settings, previewing the charge, scheduled changes, and keeping product access in sync.
General
Stripe Webhooks Without Double Charges
Stripe webhooks arrive late, twice, and out of order. How to verify, deduplicate, and queue them so billing state and product access stay in sync.
General
AI at Work Isn't a Replacement. It's a Second Pair of Hands.
A developer, a doctor, a CPA and a teacher on what they hand to AI and what they keep, what the studies add, and why the sensible time to start is now.
Retail
Computer Vision for Loss Prevention vs Shelf Monitoring: Different Systems
Loss prevention and shelf monitoring are different computer vision systems - different cameras, models, placement, data handling, and legal weight.
Retail
Combining Shelf Cameras With POS Data for Better Stockout Signals
Shelf vision plus POS data disambiguates a quiet shelf - a gap the camera sees plus zero recent sales and system stock on hand is phantom inventory.
Retail
Computer Vision for Cold-Chain and Freezer Aisles
Freezer aisles are the hardest shelf-monitoring case - glass doors, reflections, condensation, frost, and cold hardware all work against reliable detection.
Retail
Computer Vision for Queue and Checkout Monitoring
Computer vision for queue monitoring counts people waiting and estimates wait time to trigger open-another-till alerts - counting, not identifying.