Work

What is running, figures and failures included

Systems designed and run every day on the operations of NES Studio and its founder. What broke is shown too: it is the most useful part for you.

Media Factory: an automated newsroom

The need: Keep four independent media titles (EmotioNES, MindNES, KindNES, ParisianEs) alive without a salaried editorial team.
What was built: A complete pipeline: monitoring, choice of angles, three-pass drafting, visuals, human approval on Telegram, web publication. Triggered every morning, idle the rest of the time.
What production taught: On 30 August, four safeguards were added, one for each failure that actually occurred: a simulated fallback plan can no longer be published; a topic with no substance is dropped before its drafting is paid for; a document search that stayed silent for 81 days behind “0 errors” now counts as an error.

The need: Develop and maintain a dozen projects single-handed, without spending the day copying instructions from one tool to another.
What was built: A conductor designs and dispatches missions; a development team (CCO) writes code in isolated working copies; a system administrator (ClawdIA), controllable from Telegram, deploys and monitors. Each model is chosen for the mission: fast and low-cost to check, more powerful for sensitive code.
The rule that came out of it: An agent's report is never taken as proof. Agents overstate their results; the actual change to the code is checked before anything is released. Two missions reported as finished had committed without publishing anything: the completion criterion now requires a verified release.

The role: A self-hosted system administrator agent that deploys, backs up and monitors the servers, with a working memory that persists between sessions.
First incident: An undocumented script wrote into the agent's memory every night, without checking anything, that all was well: 51 identical blocks, more than half of its long-term memory. Rule adopted: only the agent writes its own memory, and no statement of status without attached evidence.
Second incident: For eleven days, a check every thirty minutes verified paths on a server that no longer existed, replaying the whole conversation each time. Diagnosed, fixed at all three causes, and told in detail in the notes.

The need: Move the technology activities into a dedicated company, on infrastructure that it owns and that is sold with it, with no dependency on a personal account.
What was built: Dedicated server, containers, reverse proxy, code forge, single sign-on, secrets vault, backups. The intended state of each DNS zone is written in code and compared with reality every day by a sentinel.
What was avoided: A first script derived the intended DNS state from the DNS in place. On the day that DNS was down, it would have copied the failure and declared it compliant, every day. It was rewritten before it was ever used.

The need: Make sure no knowledge depends on the memory of a person or of a conversation: neither the founder's nor an agent's.
What was built: A single documentation base, organised by area, exposed to agents through a secured server. Each page carries its own dated log. Agents consult it before acting and propose their updates; a human approves.
The principle: Yesterday's documentation is a hypothesis, not proof. During one diagnosis, a plausible cause already written in NES Studio's own documentation was ruled out simply by lining up the timestamps.

The context: A premium used-vehicle dealer, importing from across Europe, around sixty vehicles a year and a great deal of paperwork: service books, invoices, inspections.
What was built: Automatic reading of service books and a verifiable history for each vehicle; a document vault per vehicle; an extension that displays this history directly on the listings; an agent that drafts replies to buyers and submits them for human approval before sending, driven by three keywords: stop, hand over, book an appointment.
The business software switchover: Migration of 4,400 private customers and 520 trade customers to new management software: field mapping, deduplication, splitting to fit the vendor's import constraints.

The need: A Drive built up over seven years and several businesses, where nobody could find anything any more.
What was done: A target folder structure per business, a clean-up plan approved before anything was moved, then execution by an agent driving the interface where the API did not allow files to be moved. Result: eight business folders at the root, duplicates removed, 62 backup archives (25 to 30 GB) identified for slimming down.
Why this is an AI case: It is the most common case in SMEs: no sophisticated model, but an agent that proposes first and acts second, and leaves the human to decide what gets deleted.

Most of these systems can be adapted to an SME in a few weeks, provided the first one is well chosen.