AI agent index: /llms.txtFull content index for AI agents: /llms-full.txt

Field notesfrom the studio.

What we validate before an AI feature reaches production: evaluation thresholds, document extraction, workflow fit, and the cost model most pilots skip. Written as we ship.

Working frameworksAI · Business · DesignWritten as we ship
Featured this quarter
A small pilot is nested inside broader production boundaries, with three explicit control gates on the route outward.
AI · Production readiness

AI production readiness beyond the demo

A decision lens for workflow fit, accountable operation and sustainable deployment.

6 min readRead the post →
19 posts
A set of test cases crosses an evaluation boundary; accepted cases continue while one exception remains outside the release scope.
AI
AI · Evaluation/6 min

Set Application-Specific LLM Release Thresholds

Use evals and ground-truth data to justify a scoped release, with explicit risk tolerance and conditions for intervention.

A prompt sits within context, tools and system boundaries, showing that changing one inner instruction does not change the whole architecture.
AI
AI · System design/5 min

When a Better Prompt Is Not Enough

Locate the constraint before choosing between prompt iteration and a change to the surrounding system.

Three distinct implementation paths branch from one workflow: a prompt, an adaptive model and deterministic rules.
AI
AI · Mechanism choice/5 min

Prompting, Fine-Tuning and Rules in High-Stakes Workflows

Assess authority, system properties and changeability before committing to a mechanism.

A document passes through extraction into a review frame, with an explicit correction loop before acceptance.
AI
AI · Document AI/6 min

OCR for business documents in 2026: what to validate

Separate benchmark coverage, product capabilities, and workflow evidence before changing review controls.

Document regions and linked form fields are translated into a structured record through explicit mappings.
AI
AI · Extraction/6 min

Structured extraction from digital PDFs and scans

Define document classes, output requirements and evaluation boundaries before choosing tools.

Preparation tasks converge into one handoff while professional review remains a distinct, visible responsibility.
AI
AI · Workflow automation/6 min

Where AI Can Remove Work in Professional Services

Choose workflow bottlenecks worth testing, preserve professional judgment and measure the full delivery impact.

Three nested cost layers distinguish model calls, retries and human review around a shared feature boundary.
AI
AI · Cost modelling/6 min

Cost-modelling an AI feature

Account for tokens, retries, human review and infrastructure before committing to scale.

Three nested frames — market, programme, activation — with one route crossing all three.
Business
Business · Activation strategy/4 min

Designing brand activations that repeat

Build the activation as a system with modular parts, named owners and outcome metrics before taking it to a second market.

A four-step route from portfolio through the inheritance event to a retained client.
Design
Design · Wealth UX/4 min

Designing for inherited wealth

Treat the handover of an inherited portfolio as a product problem with an explicit retention risk, not a relationship problem.

A four-quadrant map of cabin control authority with the degraded case marked.
Design
Design · Mobility UX/4 min

When the interface becomes the vehicle

Decide which driving tasks the screen may own, which stay physical, and what happens when the software is unavailable.

A four-layer measurement stack from sample to interpretation.
Design
Design · Longevity brands/4 min

Designing longevity products people can trust

Show the evidence behind a health claim, the interval around it, and what the customer should do next.

A four-step route from published content to an owned product surface.
Business
Business · Creator economy/4 min

From audience to product

Decide which recurring audience problem you will solve as a product, and what you own when the platform changes its terms.

A four-layer stack from identity through documents and payments to the consent layer.
Design
Design · Public infrastructure/4 min

Designing on public rails

Treat identity, consent and payment rails as fixed constraints, and design the explanation layer citizens actually use.

A hub of public demand fanning out to search, clinic, pharmacy and adherence.
Business
Business · Pharma marketing/4 min

When demand arrives before the prescription

Plan for patients who arrive already informed by social platforms, and decide what your channels may say before a clinician is involved.

Concentric frames of audience, access and ownership with the innermost boundary held.
Business
Business · Luxury digital/4 min

Digital luxury without dilution

Decide which signals of scarcity you will hold constant before you expand digital reach.

A hub of one event coordinating many vendors, with the coordination layer marked.
Business
Business · Wedding industry/4 min

Designing for the wedding economy

Pick one coordination failure in the wedding supply chain and build for it, rather than building another vendor directory.

A four-stage passage through a public space from approach to exit.
Design
Design · Civic UX/4 min

Designing the city as an interface

Treat wayfinding, queueing and signage as interface work with measurable outcomes and an obligation to serve everyone.

A four-layer stack from signals through models and decisions to accountability.
AI
AI · Enterprise marketing/4 min

Marketing systems that decide

Name the decisions you are delegating to a model, the evidence required to trust them, and who is accountable for the output.

Start here

Some of our best projects started with a two-line email.

Most of our work starts with a conversation. No deck required.

Start yours