editor

Ferre Mekelenkamp

Hands-on technical engineer writing about production AI systems: state, orchestration, retrieval, evals, observability, recovery, and the places AI products fail after the demo works.

Why this publication exists

The demo is not the system.

Signal & State exists because most AI work does not fail at the moment the model produces text. It fails at the seams around the model: stale context, missing state, unsafe tool calls, unobservable runs, unclear approval paths, brittle retries, and recovery logic nobody designed.

Ferre writes from the engineering layer around those seams. The goal is not to review model releases or collect prompt tricks. It is to make production failure modes visible before teams turn a promising prototype into product infrastructure.

what the writing covers

Production AI, treated as backend engineering.

Every issue is meant to leave technical leaders and senior builders with a sharper architectural question, pattern, or failure model.

01

State and memory

What the system persists, resumes, exposes, and uses to decide what happens next.

02

Execution and control

How model output becomes bounded action through tools, permissions, stop conditions, and approval gates.

03

Evals and observability

How teams reconstruct runs, mine failures, detect regressions, and decide whether behavior is actually good enough.

Selected background

Built production software across teams

Ferre has spent 10+ years building scalable Laravel applications, AI-enabled products, and practical internal systems, with public claims of 200+ deployments and 50+ projects on ferre.dev.

Worked on durable business systems

Past work includes research portals, enterprise quote-to-purchase workflows, performance-sensitive booking systems, and editorial dashboards.

Ships inside real team workflows

The operating context is practical software delivery: pull requests, reviews, deployment pipelines, issue trackers, remote teams, and code that has to keep running after handoff.

Builds AI systems with review paths

His AI systems practice starts with one repeated workflow and designs the surrounding context, tools, review, logs, and fallback behavior before pretending the agent is autonomous.

subscribe

Get one concrete production AI essay for technical leaders each week.

Architecture breakdowns for CTOs, technical founders, and engineering leaders building AI systems past the demo. No hype, no model-release recaps, no “top 10 AI tools.” Unsubscribe in one click.

$ subscribe --channel=newsletter next issue ships fri
weekly free unsubscribe anytime
your address is used only to send the newsletter. ↵ to submit.