I engineer the systems around LLMs. Not the prompts inside them.

A reasoning model leaked its chain-of-thought and the system prompt into answers

cleanAnswer strips think blocks and leading meta-paragraphs, asserted in the unit tests

Streaming shipped with token usage hardcoded to null, so the cost panel was always empty

streamUsage enabled on the client, asserted in the integration tests

The work-authorisation matcher caught “work authorization” and missed “authorized to work”

both word orders matched, both asserted in the eval suite

ScrollTrigger measured the page before webfonts laid it out, so below-fold elements spent their entrance off-screen

the motion kit awaits document.fonts.ready before it resolves

The build crashed when no embedding key was present, instead of degrading to lexical retrieval

the projection step exits cleanly with a null stress, guarded by two regression tests

Question grammar was being scored as content, so a 64-token section matching the shape of the question outranked the section that answered it

stopwords are dropped at tokenisation, including the pronouns that carry no signal in a corpus about one person, with six ranking assertions in the integration tests

The corpus claimed the agent degrades to the retrieved source when every provider fails, and only the non-streaming path did it; the path every visitor uses returned an error

the streaming path serves the top chunk unsummarised and labels the answer degraded

The streaming path applied its reasoning filter per chunk, but a token already sent cannot be retracted, so a visitor was shown the model deliberating before the answer arrived

the opening is held until a paragraph break proves it is not working, and only then released

A corpus section grew past the embedding model’s 512-token limit, so the request returned 400 and the whole site silently dropped to keyword-only retrieval

long passages are split into windows and mean-pooled, and a rejected request now names the chunk and fails loudly instead of reporting a fallback

The graph documented that conversation history was passed per request, and nothing accepted history, so every follow-up was answered as though it were the first question asked

history rides with each request from the browser tab and is folded into both the prompt and the retrieval query

Select-to-ask sent the highlighted phrase with no section and no history, so the one interaction that proves the agent is reading this page answered like a search box

the nearest section heading travels with the selection

The colour system claimed three text tones and had two: the two quiet ones measured 1.06:1 apart, where 1.00 is the same colour

the ramp is set explicitly rather than mixed off a single axis, and every step is measured in-browser

Emphasis in prose was a 6% size change at the same colour as its surroundings, invisible to the scan that hunts for bolded phrases

a second semantic accent, chosen by simulating three kinds of colour blindness rather than by eye

Entrance motion ran correctly and could not be seen: 14px over 500 ms is below the perceptual threshold for motion

30px over 620 ms, which clears the threshold without reading as a template

The site URL in the corpus pointed at an address that was not this site, feeding the canonical tag, the entity graph and the share card

corrected at the source, where every surface reads it from

Two entrance systems owned one element’s opacity: a panel faded itself in with @starting-style and was also swept into the scroll batch, so GSAP wrote an inline zero while the CSS transition was mid-flight and whichever finished second decided what the reader saw

the element that animates itself says so, and the batch skips it

The net that catches an element stranded at zero skipped anything inside [data-state], and panels set that attribute unconditionally, so every panel on the page was exempt from the sweep written to rescue it

only a running instrument is skipped, because an idle panel at zero opacity is reporting nothing

That same net ran once, a second after mount, while the stranding it guards against happens whenever a reader crosses several batch boundaries quickly, which is at any point in the session

it runs again whenever scrolling stops

None of these threw an exception. That is the point: a leaked monologue renders as text, a null token count renders as blank, and a missed pattern returns a confident answer to the wrong question.