Skip to main content
Architecture

How we designed things that are actually running.

Written for the person who will notice if we are wrong. Each piece names the constraints, the trade-off we took, and what it cost us - and most end with something you can go and check for yourself.

One minute of speech, measured three ways: 120 KB as the browser records it, 1.92 MB as the model wants it, and about 900 bytes of transcript. Only the last one is needed.
Shipped and running
23 Sept 2026

Your speech-to-text sends the recording. It only needed the words.

A speech feature exists to turn a voice into a paragraph. Most of them upload the voice to a vendor to get it. The design where the model runs on your own server, what the queue of one costs, and the part nobody admits - the audio is still stored, on purpose.

ffmpeg -i clip.webm -ar 16000 -ac 1 -c:a pcm_s16le clip.wav && ls -l clip.*Read it
Two designs side by side: a model with a send tool reaching an inbox directly, and a model that only returns a draft to an editor, an audit and a queue.
Shipped and running
12 Sept 2026

Let the AI write your outreach email. Never let it send it.

A model that can send will one day send the wrong thing. The design that lets a model draft into an editor, runs every draft through a rulebook in code - nine blocks, fourteen warnings - and keeps the send button for a person. Every rule quoted from a system in production.

Ask the vendor: which function sends, and can the model call it?Read it
A context budget filling up: rules and retrieved documents competing for the same window
Shipped and running
10 Sept 2026

Why your RAG chatbot ignores the rules you gave it

A knowledge base outgrows the context budget, retrieval starts choosing what the model sees, and the rules you wrote become documents competing for space. How to measure what your bot actually receives, and why some of it must never be scored.

context budget ÷ knowledge base sizeRead it
The wildcard rule /website-development/:slug* alongside the live page it caught, and the negative-lookahead fix
Shipped and running
6 Sept 2026

Your redirect list is about to swallow a page that still works

A wildcard redirect does not know which of your pages are still live. The rule that ate a working archive, the exclusion that fixed it, and 178 rules anyone can test against this domain with curl.

curl -sI https://boffincoders.com/website-development/projectsRead it
A terminal showing the sitemap diff: 10,722 URLs before, 10,723 after, nothing dropped
Shipped and running
6 Sept 2026

Your rebuild will drop URLs. Google will tell you in six weeks.

We moved a 10,723-page site to a new framework. The sitemap diff that proved no URL moved, the four things it cannot catch, and the outage that mattered more than any of it.

https://timezones.in/sitemap.xmlRead it
bundletool output for the PDF Toolkit release with android.permission.INTERNET highlighted as added by a dependency
Shipped and running
4 Sept 2026

Your Android app asks for internet permission. You never wrote it.

The manifest you write is not the one that ships. Every library brings its own, the build merges them, and a permission can arrive four levels down. How to see what your app really asks for, and the one line that refuses it - from an app that had told people it needed no internet.

bundletool dump manifest --bundle=app-release.aabRead it