Putting an LLM in production without giving it autonomy
Where the confidence threshold goes, what human review has to catch, and which decisions a language model should never be allowed to own in production.
ReadInsights
What we have learned shipping custom software, migrating legacy systems, and putting AI to work where it earns its keep.
Where the confidence threshold goes, what human review has to catch, and which decisions a language model should never be allowed to own in production.
ReadHow to replace a legacy system in slices while it keeps running, which slice to take first, and the three conditions that make the pattern fail in practice.
A discovery that produces a proposal is a sales process. Five artifacts that make it an engineering deliverable you own and can hand to any vendor.
An anticorruption layer lets new systems talk to an untouchable ERP without inheriting its data model. What belongs in it, and how to keep it thin.
Screen automation is fast to build and expensive to keep. A decision table for choosing between RPA and a real integration, and how to survive RPA.
Off-the-shelf is cheaper until you count configuration, integration, and the processes you bend to fit it. A test for deciding which side a capability falls on.
A framework for estimating custom software cost from the factors that actually move it, so you can challenge a proposal instead of comparing headline totals.
ELT is the default advice and it is not always right. Where each pattern belongs, what the warehouse bill hides, and the constraint that decides it for you.
Case studies and team bios do not distinguish vendors. Six questions that surface real engineering depth, and the answers that should end the conversation.
A signed BAA is not a compliance certificate. What it commits your software vendor to, what it does not cover, and the questions to ask before signing one.
The Security Rule technical safeguards written as engineering decisions: what each one requires, what satisfies it, and where teams usually fall short.
Hourly rate is the smallest term in delivery cost. What time zone overlap, rework, and context loss actually cost, and when each model is the right choice.