I am going to publish how YuktiBOS is doing while we build it, including what is broken. This post explains why, and shows where things stand on 4 to 5 November 2026.
Why publish stats while we build
Three reasons, in order of how much I believe them.
One: it keeps us honest. When I wrote that our automated judge agreed with careful readers at a correlation of 0.27, I could not unknow it. Numbers written down in public are harder to explain away. The series of posts before this one is built from our own working documents, with the numbers left in.
Two: AI-built software has a credibility problem and the cure is evidence. The honest account of an AI-built system includes the bug list, the wrong fixes and the bill that did not reconcile. Our Content Lab bug list has 25 items. I would rather you read it than guess.
Three: it gets us better inputs. We are one founder and a small team building a lot. The biggest failure in our content tests came from a wrong line in our own source material, and I found it only because a reader spotted a pattern. Outside eyes help.
I am not claiming results we do not have. Quality is below my own bar. A piece I would publish under my name with no edits is my 9.5 out of 10, and we have not reached it.
What YuktiBOS is
YuktiBOS is an AI business operating system. Yukti is Sanskrit for strategy. It runs an agency's work: client onboarding, strategy, breaking work into tasks, assigning, reviewing and approving, plus internal operations and a CRM. About half the work is done by AI agents, and a person can take over at any step.
Two beliefs shaped it:
- Everything AI can do end to end, a person must also be able to do, with the same structure and tools. With no AI key at all, the process still works. People write, review and post.
- Clean, not lean. Show only what is needed now, with the next step nearby. Never cut functionality.
Where it stands
As of this snapshot, the main branch is published and migrations are applied through number 0287. The feature index lists 53 features, each with a help article and a spoken explainer script. The state of the main parts, as I can verify from our notes:
| Area | State |
|---|---|
| Content (posts, articles, calendar, publishing, results) | Built and live. Quality below target |
| Content Lab (compare models and methods) | Built, live. Scoring v8 built, calibration result pending |
| Fact store, never-say gate, client brief | Built and live |
| Article workbench: seven stages, a person can take any stage | Built and live |
| Human-first mode ("I'll write it", writer, reviewer and poster roles) | Partly built: the article workbench hand-over and "Mark as posted" exist; the full mode is next on the list; help article written |
| Directional Reviewer (weekly direction at planning) | Rebuilt; it now steers at planning |
| Market Finder, story cards, strategy calendar | Built |
| SEO v1 (audit, change list, re-check, AI-search and Search Console panels) | Built; live check on iipta.com still to do |
| Client tickets: auto-start, one gate | Built |
| Auto-posting to LinkedIn and Instagram | Built; needs the accounts connected |
| CRM, Reports, Tools (Voice, Brand, Market Finder, Offer, Pricing) | Present; no status note this week |
| Inbox merge, full ticket pipeline, SEO Checkup public page, own email system | Not built yet |
| IIPTA employee readiness and role dashboards | Spec and critique done; build next |
What we have learned so far, in five lines
- Rules do not make a good writer. We had cut 65,000 to 75,000 characters of instructions and the score moved the most.
- Plain chat is fluent and generic. Orchestration beats it by a wide margin, mostly by stopping it inventing.
- An automated judge agreed with careful readers at r = 0.27. We rebuilt it.
- Raw knowledge search did not earn its place. A checked fact store did.
- One of our biggest failures was a wrong fact about my own company, fed to every model.
Each of these has its own post.
What is open
- Quality is not at 9.5.
- The scorer rebuild has no published result yet.
- Spend figures are over-reported by about 23% in the past; the rupee rate and price table are unchecked.
- The human read was by an AI editor. A person should read a sample.
- Employee-facing features need a privacy and fairness review before I describe them publicly.
Come and try it, or tell us what to build
If you run a small agency, a consultancy, an institute or a founder-led business, I want to hear from you.
- Feature requests. Tell me the job you want done. I will say whether it exists, is half built or is not started.
- Trials. I will onboard a small number of businesses for a trial. Your data will stay yours. I will tell you what we log and where AI is used.
- Challenges. If you think one of our numbers is wrong, say so. Several of them have been wrong.
Contact: request a trial or a feature.
Numbers at a glance
| Item | Value |
|---|---|
| Features indexed, with help article and script | 53 |
| Database migrations applied | through 0287 |
| Content Lab pieces in the key / read blind | 363 / 284 |
| Human vs automated judge agreement | r = 0.27 |
| Content Lab bug list items | 25 (22 fixed or closed, 3 need my decision) |
| Competitor SEO products profiled | 29 |
| Real benchmark posts collected | 20 + 30 + 36 + 64 brand posts (plus 13 for IIPTA) |
| Target for judge agreement | 0.6 or better |
We are onboarding a small number of businesses for early access, and we read every feature request.
