Point of View · Field Notes from Forward-Deployed Engineers
Field notes, working diagrams and patterns from twenty-one programs. Written by the people who shipped them. No vendor framework. No prophecy. No hedging.
We publish three things: field notes from work in flight, boundary diagrams from programs that shipped, and the occasional manifesto when the industry needs one. We do not publish marketing.
36
Pieces published · Last 18 months
Manifesto · 12 min read · Mar 2026
Featured
Three years of "AI initiative" budgets, two years of "AI strategy" engagements, and the picture is the same in every customer we walk into: pilots that demo and never pass audit. The fix is not another model. It is treating the work as a workflow that happens to have a model in it.
Read the manifesto →Series · 9 Pieces
Redacted, audit-signed boundary diagrams from shipped programs. The artefact, the rationale, and the conversation that produced it. Read the diagram an FSI compliance officer signed in March, in 6 minutes.
Browse the series →Series · 8 Pieces
The recurring architectural patterns that did the work. EDA smart queues. Babelfish-cutover discipline. Bedrock Agents with scoped reversible tools. Each pattern with the program it shipped in.
Browse the series →Series · 5 Pieces
Longer pieces written four times a year on what is happening to the AWS modernization and agentic-AI market. Specific, dated, falsifiable. No category creation. No future-of-anything.
Browse the series →Patterns we noticed on the way to shipping · 14 pieces
A one-page artefact, signed by Internal Audit in week two, is what makes the next ten weeks linear. The diagram names where PHI travels, where models are invoked, and where a human has to be on the loop.
If a regression on the audit-citation gate can ship to production, the harness is decoration. We treat eval as a CI step that blocks, not a dashboard that observes. Aedeon, the way we use it, has rejected releases.
Database migrations live or die on what you can prove the morning after cutover. Production-traffic replay, with diff-on-result, is the artefact the finance committee actually reads. Synthetic load tests do not survive that meeting.
EDA clusters built for peak capacity waste money and time-to-tape-out. Smart queues that reshape capacity against the actual job mix change the economics — Synaptics saw 40% throughput and 75% lower queue wait.
Chat windows are the wrong UI for production agent work. Worklists are the right one. The agent prepares; the reviewer disposes; the policy paragraphs the recommendation rests on are cited inline.
SQL Server retirement that does not require rewriting the apps is, in this market, a Tier-1 economic event. Per-plant cutovers, replay validation, licence window closed on its own date. Not glamorous. Highly profitable.
Five published · The sixth is in flight
A close look at what the four big database vendors are calling "AI-ready" in their 2026 product collateral, and what an actually-agent-ready substrate has to do that none of them quite do yet.
An argument, with named programs, for why the strategy deck is now the thing slowing AI adoption down. The unit of work is the agent that passes audit. Everything else is a tax on the customer's calendar.
What we mean by "forward-deployed", what it costs to staff this way, and why we believe it is the only honest way to ship a non-trivial AWS workload in 2026. With three programs as the evidence.
A re-framing of what AWS shipped with Bedrock, why the audit posture matters more than the model menu, and what a regulated-industry program team should be measuring at the platform layer.
A long piece every quarter from our engineers. Plus the field notes worth surfacing that month. No webinar invitations. No event funnels. Cancel-by-reply.