Writing
What I am working on, and what it taught me.
I write in two streams. One is about operating and positioning a business, what actually drives it and how to say so. The other is about the systems I build, how they work and where they break.
The two streams are drawn below. The essay sits where they meet, and the pointer stirs the water.
Operating and positioning a business
Public, hybrid, or private? A field guide to where your AI should run
Model choice isn't a one-time decision — it's a trajectory. A practical way to decide where each workload belongs, and when to move it.
Why automations fail — and how to make them stick
The failure mode of automation isn't technical. It's that nobody uses it. Designing for adoption from the first conversation.
The systems I build, how they work and where they break
How do you qualify 'credible'?
Saying influence should be credible is easy. Turning that into a repeatable selection process — research, a tooling stack per tier, and named criteria every candidate must pass — is the part that makes it an operation rather than an opinion.
Do the matching by hand until the algorithm has something to learn from
A talent marketplace I co-authored at Xynteo sequenced its product roadmap so that the matching was manual first, calibrated by people second, and automatic last. That order is not timidity. It is the only way the algorithm gets a training set worth having.
Influence is a measurement problem, not a casting problem
Most influencer work picks people by audience size, campaign by campaign, and starts again each time. The product I co-created at Publicis treated it as a standing instrument instead: monthly reports on content, on who moves it, and a required answer to what anyone should now do about it.
Credibility and reach are not the same purchase
Influence has two independent properties and most buyers pay for one while assuming the other comes free. The model I co-created at Publicis separated them, ranked five kinds of influence by what each actually supplies, and put the measurement on the response rather than on the person.
A narrative you have to police is not working
Companies still treat their story as something to be controlled: written tightly, distributed carefully, corrected when it drifts. The method I co-authored at Xynteo assumed the opposite — that a narrative only functions once other people can make it their own — and it had a vocabulary precise enough to run on.
An error handler that retries what's safe and escalates the rest
Every workflow failure on our box used to produce the same Telegram alert, whether it was a five-second network blip or a real bug. A local model now classifies the failure first, and only a transient one on an explicitly allowlisted workflow gets retried.
One workflow, many clients
The usual way to serve a second automation client is to copy the first client's workflow and edit the copy. We built the alternative before we needed it: a registry row per tenant, and a sub-workflow that resolves who is calling before anything else runs.
Three workflows for a DTC brand's first week
DTC and CPG e-commerce is the vertical we hear the most automation demand about and find the least operational detail written about. We built three small workflows sized for a first engagement week, and reported which parts run live today, which run real analysis over fixture input, and which are wired templates waiting on a client credential.
The DSAR pipeline that stops at a person
A data subject access request asks a company to compile one person's personal data in one place. We built the pipeline that handles GDPR and CCPA requests with the private model doing the sensitive step and a human sending the response.
Measuring 'public where it makes sense, private where it matters'
Our positioning is a claim about the tradeoffs between public and private models. This week we built a harness that measures it: the same 10 tasks, run three ways, scored for accuracy, latency, cost, and what leaves the building.
A visibility check that reports what six questions on one engine can actually show
Buyers now put questions to Claude, ChatGPT and Perplexity that they used to type into a search box. We built a free tool that asks Claude six real buyer questions about a brand, with live web search, and emails back what it said: two counts over a stated denominator, and no invented score.
We automated ourselves first
Before selling automation to anyone else, we built sixteen n8n workflows that run our own firm — on one small server, with a 0.9% failure rate. Here's what runs, what it costs, and the design rules that made it stick.
The deployment path, as running code
Our positioning is 'public where it makes sense, private where it matters.' This week we turned it into two working systems: a PII-aware model router and a RAG assistant that answers from our own hardware. With receipts.