Research / Engineering, examined
Ideas worth putting to work.
Practical writing on building AI systems, choosing the right architecture and bringing people into the work. Research informs the choices; the business gives them a purpose.
An essay from the worktableFeatured perspective / People & delivery
Forward-deployed engineering: build the system with the people who will use it
A team can have access to capable AI tools and still struggle to use them well. The missing work is often close to the business: choosing a suitable task, recognising a weak answer and knowing what to do when the usual path fails.
Read article6 min readThe research library
01 — 0801People & delivery
Forward-deployed engineering: build the system with the people who will use it
How embedded engineering, workflow-specific AI training and practical handover help a team take ownership of new systems.
An essay from the worktable6 min read
02Architecture & delivery
What a production AI system needs around the model
A practical architecture review for turning an AI feature into a reliable service, from permissions and data to operations and ownership.
Architecture notebook5 min read
03AI & security
Design an AI agent around the actions it is allowed to take
How to check an AI agent’s permissions, what a reviewer needs to see before approving an action and what records to keep for review.
A security field note4 min read
04Quality & evaluation
An AI evaluation set that helps you make a release decision
Build useful AI evaluations from representative work, explicit failure criteria and evidence that connects a result to a release decision.
A practical guide4 min read
05Knowledge & retrieval
Document access in an AI assistant
Design a useful RAG system with source ownership, permissions, document versions and evidence that a reader can inspect.
A technical essay4 min read
06Economics & operations
What an AI result costs after review and acceptance
A practical way to compare AI operating cost, review effort, latency and quality without optimising one number at the expense of the workflow.
A short working note4 min read
07Vision & applied ML
Testing computer vision in real conditions
How to evaluate computer vision under real camera conditions, select useful thresholds and design human review around costly mistakes.
A field scenario5 min read
08Data & reliability
The response that never came back
A fictional integration incident about uncertain outcomes, stable identities and the person who has to establish what happened.
A fictional incident4 min read