Insights
Engineering writing, not thought leadership.
Things we have learned doing the work, written for people who will have to do it themselves.
18 June 2026 · 4 min · QAI Labs engineering
Sizing GPUs for a 70B model you have to host yourself
The arithmetic that decides whether your cluster survives Monday morning (weights, KV cache, headroom) and the three assumptions that most often make it wrong.
6 May 2026 · 3 min · QAI Labs engineering
MCP and A2A solve different problems, and most teams need one of them
Two protocols get compared as if they were alternatives. They sit at different layers, and the honest answer for most systems is that you need the tool layer and not the delegation layer.
9 April 2026 · 4 min · QAI Labs engineering
Why we make agent control flow a graph, not a conversation
Emergent orchestration is the fastest way to a working prototype and the slowest way to something a supervised process can adopt. What we do instead, and where we still let the model decide.
24 March 2026 · 3 min · QAI Labs engineering
Evaluating agents when there is no single right answer
Accuracy works when there is a correct label. Most agent work has no label, several acceptable answers, and a trajectory that matters as much as the output. What we measure instead.
11 February 2026 · 3 min · QAI Labs engineering
Getting model weights into an air-gapped enclave without breaking accreditation
The file transfer is the easy part. The hard part is proving, eighteen months later, exactly which artefact ran on which day, and building a path that a security team will approve twice.
15 January 2026 · 3 min · QAI Labs engineering
The handover problem, or what "done" means on an AI engagement
A system that only its builders can operate is not finished. What we hand over, how we test that the handover worked, and why we write the exit plan at the start.
Start with the constraint.
Most of these projects are shaped by what you cannot do rather than what you want. Data that cannot leave the estate, a model you cannot host with a third party, a decision somebody has to justify to a regulator. Tell us yours and we will say honestly whether we can work inside it.