Agentic AI Design
Design and build agents that use tools and work across systems, with clear permissions and human approval points.
AI systems with clear scope and measurable performance
B12Y helps organisations design and build AI systems, measure their quality, and decide what to improve, replace, or retire. Based in Auckland, we work internationally on AI and complex technical systems.
Services
Assess, design, build, and evaluate systems
We work with organisations to select, build, evaluate, and operate systems that fit their requirements.
AI-First Transformation
Identify where AI fits into your organisation's work, then plan the changes to workflows, responsibilities, and controls.
AI Rationalisation
Independent assessment of AI systems against operating cost, review effort, reliability, risk, and the alternatives available.
AI Quality & Evaluation
Evaluation cases, regression suites, and production measurement that make system changes visible and support reliable release decisions.
Common reasons to engage
Start with a system or decision that needs to change
B12Y can enter at assessment, design, implementation, or evaluation when the next technical decision is unclear.
AI experiments are not reaching production.
Quality cannot be measured reliably.
An agent needs controlled access to tools and systems.
Operating cost or review effort no longer justifies the system.
How an engagement works
From assessment to implementation
The scope and evidence required at each stage are agreed before implementation proceeds.
01
Diagnose
Establish the current system, the constraint, and the outcome that matters.
02
Decide
Determine whether to build, change, replace, or retire the capability.
03
Deliver
Implement and measure a system the client team can operate.
Recent technical notes
Practical writing about AI systems
Decision frameworks, implementation details, and operating considerations from current areas of work.
01Design
Recovering an agent after a runtime failure
A recovery design for saving useful agent state, checking unfinished work, and controlling the return to operation.
Read article02Optimise
Choosing a quantisation for your local model
Budget runtime memory, compare candidate files, and test whether a smaller representation still supports the workload.
Read article03Secure
Testing what an agent can reach and change
Use disposable fixtures, enforced access controls, and independent records to test an agent's capabilities.
Read article