Reference Reviewed Sep 2026 Vendor-neutral

AI reference architectures.

Published designs for the systems we build most often — the components, the failure modes, and the numbers you should hold a vendor to. Written to be useful whether or not you hire us.

WHY THESE EXIST

Most AI case studies are unfalsifiable.
These are checkable.

An agency case study typically reports a percentage and a happy client. You cannot verify the percentage, you cannot see the architecture, and you cannot tell whether the result would transfer to your problem. It is marketing shaped like evidence.

A reference architecture is the opposite trade. It shows the actual component graph, names the models and infrastructure, states where the design breaks down, and lists what you should measure before you believe anyone — including us. If the design is wrong, an engineer reading it can tell.

These are designs we work from, documented in the abstract. They are not client reports, and none of the figures on these pages are claimed client outcomes — they are design targets and published benchmarks, labelled as such.

HOW TO READ THESE

PROBLEM: What breaks, and why the obvious fix does not work
DESIGN: The component graph, with the models named
FAILURE MODES: Where it degrades, and the mitigation
WHAT TO MEASURE: The evaluation that tells you it works

THE POSITION

A percentage you cannot verify is marketing shaped like evidence. A design an engineer can argue with is the real thing.

THE ARCHITECTURES

Document intelligence over regulated archives

Healthcare · Legal · Insurance

Extracting structured fields from tens of thousands of long, unstructured PDFs without sending a page to a public endpoint — and making every extracted value open to the exact sentence it came from.

Air-gapped inference · constrained decoding · span-level citation

→

Support deflection that knows when to stop

SaaS · E-commerce

A retrieval agent over product docs and account state that answers what it can evidence and hands off everything else — with the handoff threshold as an explicit, tunable parameter rather than an emergent accident.

Hybrid retrieval · confidence gating · human handoff

→

Risk and compliance agents with an audit trail

Financial services

Agents that read policy, apply it to a case, and produce a decision a regulator can reconstruct — every tool call, retrieved passage and intermediate judgement persisted as an immutable record.

Deterministic tool use · full-trace logging · four-eyes review

→

High-throughput document pipelines

Logistics · Supply chain

Multimodal extraction at volume, where the interesting problem is not accuracy on a good scan but throughput economics and what happens to the small fraction of documents that are damaged, rotated or in an unexpected language.

Batch inference · exception routing · cost per document

Keep reading

Evaluating AI Systems

→