Case studies

How sharing a project chat could expose private context documents through AI responses that quoted or paraphrased them. Fixed with a pre-share similarity scan with user-controlled redaction, a context disclosure notice on all shared project chats so recipients understood the AI had access to private material, and revocable share links with optional expiry dates.

How raw volume thresholds were flagging paying power users while missing sophisticated abusers who distributed their load. Fixed with multi-signal behavioural scoring covering timing regularity, session entropy, and fingerprint diversity, cross-account soft identifier matching for free-tier cycling, and a graduated response of throttle, verify, suspend, deactivate rather than binary blocking.

How limit checks run after streaming started were causing responses to cut off mid-sentence with no explanation. Fixed with pre-query token reservation using atomic compare-and-reserve, a credit system with per-model cost weights, progressive usage awareness at 75/90/100 percent, and usage-pattern-based upgrade recommendations.

How raw conversation history hit token limits within weeks and naive truncation was removing the most important early context first. Fixed with a three-tier architecture (permanent core, rolling importance-weighted summary, live window), query-adaptive tier selection that only sent the tiers each query actually needed, and asynchronous summarisation so it never added latency.

How concurrent messages from multiple users were corrupting the shared context and how one talkative participant could fill the context window for everyone else. Fixed with a per-session ordered message queue using atomic sequence numbers, batching of rapid messages into single completion calls, split context budgets per participant, and named attribution in the context so the AI could address people individually.

How keyword-based routing was misclassifying mixed queries and adding 800ms+ of visible latency. Fixed with a feature-based lightweight classifier running in under 30ms, parallel model warm-up to hide the classification overhead, a capability-weighted scoring system that balanced complexity and cost, and a user feedback loop feeding weekly retraining.

Explains the admin system built for managing a live B2B agent network, not just a product catalogue. Covers agent onboarding approval queues, per-agent portfolio visibility, scheduled pricing updates with effective-date control, platform-wide order intelligence that flags credit breaches and duplicate orders asynchronously, read-only cart and wishlist visibility for support investigations, and session management with remote termination for security incidents. Designed explicitly for non-technical operations…

Documents the deferred B2B payment system where orders are placed on credit and settled later. Covers the order-level payment state machine (Unpaid, Partially Paid, Paid, Overdue, Disputed), per-customer balance summaries with credit limit monitoring, manual payment recording with oldest-first allocation logic, an immutable payment audit trail with reversal entries, and automated overdue alerts that turn collection follow-up into a structured daily workflow rather than a memory-dependent task.

Details the per-customer persistent cart model that lets agents accumulate order drafts across multiple shop visits before submitting. Covers the bulk cart building interface, wishlist-to-cart conversion, the last-order-as-template workflow that speeds up repeat ordering for established customers, atomic order submission to prevent partial states, and a local-first cart update approach for agents working on poor mobile connections.

Covers the B2B-specific account structure where the agent, not the end consumer, is the primary actor. Explains the approval-gated registration flow, how each agent manages a portfolio of customer shop accounts, how session context is scoped per customer to prevent cross-account errors, and how the wishlist was reimagined as a field visit journal for capturing shop requirements before formally placing orders. Mobile-first design considerations for field use are also detailed.

Documents the full operational control surface used by non-technical platform operators. Covers product and catalogue management with CSV bulk import, real-time order tracking with a complete event log, cart and wishlist visibility for customer support, payment gateway oversight across multiple providers, and user account management with session-level visibility and remote termination. Built on role-based access control with an append-only audit log covering every admin action.

Details the cart and checkout infrastructure designed specifically for high-volume orders with 50+ line items and large unit quantities. Covers real-time tiered pricing computation on every cart mutation, wishlist-to-cart bulk conversion, early stock validation before checkout, CSV order import for repeat buyers, atomic inventory reservation to prevent overselling under concurrent load, and a multi-stage checkout flow for large-value transactions.