About This Casebook
Engineering decisions don't live in the code — they live in the reasoning behind it. This casebook documents the architecture choices, failure mode analysis, trade-off records, and operational patterns behind systems I've designed and built. Where source is proprietary, the design and reasoning is public.
Control-Plane / Data-Plane Separation in LLM-Driven ERP
How to build an agent-driven ERP middleware that stays reliable when the model is unreliable. Covers provider adapter boundaries, read-model vs. write-path trade-offs, and decision records for key architectural choices.
Camera-to-Event Pipeline for Constrained Hardware
Architecture for running computer vision on Raspberry Pi and Kneron KL520 class devices with 512 MB–4 GB RAM, no GPU, and intermittent connectivity. Covers device budget, offline relay strategy, model/resource trade-offs, and rollout plan.
Observe → Plan → Validate → Approve → Execute State Machine
Threat model, approval requirements, and replayable trace design for browser automation agents. Covers prompt injection boundaries, validation gates, and the reasoning behind requiring human approval for destructive actions.
Engineering Principles
Validate before you run. // Dry-run mode, policy checks, schema contracts Bound your queues. // No unbounded growth; explicit drop policies Measure before you claim. // Benchmarks with methodology, not marketing Name your failure modes. // Threat models, offline strategies, drop behaviors Separate control from data. // Clean boundaries between orchestration and execution