Development snapshot, Nov 2026: what YuktiBOS is today
What YuktiBOS is today, what is built, what is still open, and an invitation.
Articles
Research, product and strategy notes, with the numbers left in.
AI-built software has a credibility problem, and the cure is evidence. These articles come from our own working documents: the bug list, the wrong fixes and the bill that did not reconcile.
All articles
23 of 23 articles
Why we built a Content Lab
Our AI posts cost Rs 5 and were “pathetic”. Plain chat beat our pipeline, so we built a lab to stop guessing.
284 pieces read blind: our automated judge agreed only weakly (r = 0.27)
We checked our own scorer against careful readers, and it mostly failed.
We fed our AI raw knowledge, and truth got worse. So we moved to checked facts.
More context gave worse posts. We replaced raw chunks with a checked fact store.
The low-budget instruction: two instruments, two answers
“Be efficient” looked harmful to the scorer and harmless to readers. What that says about measuring.
We collected real posts to learn what "good" looks like. Here is what the data can and cannot say.
We collected real posts instead of guessing what good looks like. Here is what the data can and cannot say.
Cost engineering: what we learned, and the bill we could not reconcile
Cost is a system-design problem. We traced every call, and our own spend figures were still wrong.
The board game that wasn't: a hidden framing in our own inputs
91 of 284 pieces called a company a board game. The cause was our input, not the models.
We read 29 SEO products before building ours. Here is the gap we chose.
29 products read, six groups, one gap: the full loop with a human approval, in plain language.
The voice tool: how we stopped asking people to describe their voice
We watched what people do, gave them a finished document, and used speech for ideas, not style.
The brand tool: beliefs can't be written by AI, so we surface them
Quotes checked by code, six parts, no colours. AI can write, but beliefs it cannot write.
Market Finder: how to pick a market when you don't know the basis
Code ranks, AI drafts, a person picks. Plus stickiness and story cards.
The Offer and Pricing tools: let the AI suggest, let the person pick, let code do the arithmetic
Arithmetic from typed numbers only, an offer card, and an honest note on what is not yet reviewed.
The Content Lab as a tool: what it does, what it cost, and the bugs
Estimate first, budgets per run, and a 25-item bug list.
The article workbench: seven stages, and a person can take any of them
Stages outside the engine, fact ids, and a hand-over a person can take at any step.
SEO in YuktiBOS: audit, change list, approval, re-check (and the public checkup we haven't built)
Find, word, approve, apply, re-check, and the public checkup we have not built yet.
Standalone tools: the plan to let friends use the strategy tools without a client
Seven leaking lists, uncounted AI spend, no new role: what is built and what is only a plan.
We graded 17 AI deliverables. Two got a B-minus.
Everything went through one machine, and “Done” did not mean done.
The Directional Reviewer: from "read everything" to "write one line"
Three redesigns in five days, driven by the lab.
From "lean" to "clean, not lean"
Lean is what you build. Clean is what the user sees.
Publishing on Replit: five ways a deploy surprised us
A schema diff that runs first, and a database driver that landed in the browser bundle.
The Rs 100 cap that became Rs 10,00,000
Two multiplications, a refused bandage, and the sequel in November.
Critique before you build: what a second reviewer found in our own specs
Six of twelve build steps in one spec rebuilt things that already existed.
The judge that cost Rs 21 a piece, and a correlation I don't trust yet
An offline correlation of 0.92 is not a result.
Think one of our numbers is wrong?
Tell us. Several of them have been wrong. Or tell us the job you want done and we will say whether it exists, is half built or is not started.