A demo and a Tuesday are different kinds of things.

An assistant I built kept returning confident, generic answers for days after its model API had failed. Nothing alerted. The uptime check passed, the endpoint returned 200, the logs were clean. I caught it by reading an answer against what should have been retrieved and noticing the passage was not in there.

Whether a process ran and whether its output was correct are two different questions. Only one of them was being asked.

• So the guard came first, not last. When retrieval returns nothing usable, the assistant has to say so. It is not allowed to fill the gap.
• And the same rule went into the document pipeline. A rule engine reads the finished file and raises an error rather than shipping something that breaks its own contract.
• And into the data layer. Any write that would shrink the record is refused. That one exists because an unguarded write truncated a file and cost weeks.

A system that fails loudly is cheaper than one that fails quietly. The quiet one bills you for months before anybody notices.

Shelf one · AI systems in production

Retrieval, and the cost of running it

The assistant above, running on a live WordPress site. Content is chunked into a full-text index and the top passages are selected per message, weighted so core material outranks incidental pages.

• Retrieval, not context stuffing. Five passages per question out of 735. The model never sees the whole site, which is what keeps the answer traceable.
• A spend cap that closes the endpoint. Per-user rate limiting on top. A demo does not need either. A thing that runs on a Tuesday does.
• Every interaction logged, and the schema versioned so a change migrates forward instead of breaking on deploy.
• Reindexing is incremental. Saving one article rebuilds one chunk. Fixing a typo does not rebuild the knowledge base.

Shelf two · automation with a brake on it

An agent-operated document pipeline

Sources live requisitions from applicant tracking system APIs, screens each against hard eligibility rules before any work happens, and generates a tailored document set for the ones that survive.

• It reads the source, not the aggregator. In one sweep, 168 listings passed a text screen. Six were genuinely what the card claimed. The rest misreported pay or location. Reading the employer's own API is now a rule rather than a preference.
• Screening runs before generation, so the expensive step never runs on something already disqualified.
• The output is inspected, not trusted. The rule engine has refused to ship more documents than I would like to admit.

Shelf three · the translation layer

Explaining AI to people who were never invited into the conversation

Forty-four articles on mstimaj.com, each shipping with an interactive machine the reader can break: change the input, watch the output move, keep the intuition afterwards.

• A tokeniser you can type into, which splits your sentence into chips and tells you plainly that it is a simplified split and not real BPE.
• A cost calculator that takes a token count and extrapolates the monthly bill, because the number people actually need is the one on the invoice.
• A retrieval toggle. Same model, same question, retrieval off and then on. Off, it hedges and nothing marks it as a guess. On, it cites.

Shelf four · one post, several platforms

Cross-posting without doing the work five times

Publer wired into the Meta, X, and YouTube APIs, so one piece of content is adapted and scheduled across platforms from a single source instead of retyped per app.

• One source, several shapes. The same content gets reformatted per platform's real constraints, not just resized.
• Scheduling is the point. Posting on time matters more than posting from memory, so the queue runs on its own.