Notes from building six products at once.

What makes a workflow ready for a multi-agent pipeline

Not every process is a good candidate for automation by AI agents — and getting this wrong is the fastest way to ship something that looks impressive in a demo and falls apart in production. When we designed Anmolytics' 12-agent workflow, the first filter wasn't "can an agent do this step," it was "can we define a clear, checkable output for this step." Topic discovery, drafting, on-page SEO and publishing all have outputs a reviewer can evaluate quickly — a headline either targets the right intent or it doesn't. Steps without a checkable output are where we keep a human fully in the loop rather than trying to automate around ambiguity.

The second filter is recoverability. A 12-step pipeline will produce a bad output somewhere eventually — the question is whether that failure is caught before it reaches a reader. Anmolytics' team-based approval layer exists precisely because performance monitoring at the end of the pipeline is a feedback signal, not a safety net; by the time an underperforming article shows up in analytics, it has already been live for days. So every niche deployment gets a review checkpoint before publishing, not just after.

The pattern generalises beyond content. Tutorspire's AI Study Planner works the same way: it proposes a day-by-day schedule (checkable, adjustable output), then adapts based on whether sessions were actually completed (a recoverable feedback loop), rather than trying to predict a student's whole semester perfectly upfront.

Why we build one platform per audience, not one platform for everyone

It would be simpler, on paper, to fold Eduledge's consultancy model and Global Study Compass's free self-serve guides into a single product with a paywall in the middle. We've deliberately kept them separate, because they serve two different relationships to the same problem. Someone using Global Study Compass wants to do the work themselves — a checklist, a guide, a quick answer from the AI assistant — and anything that feels like a sales funnel breaks that trust immediately. Someone contacting Eduledge wants a person to own the outcome for them, end to end, including the parts that don't have a clean self-serve answer, like negotiating an education loan.

Trying to serve both needs from one interface usually means serving neither well. So instead, each Vuray platform is scoped tightly to one audience and one relationship to the problem — and where it makes sense, like routing a Global Study Compass visitor toward Eduledge for hands-on help, or an Eduledge student toward Vuray Travels for their flight and visa, the products point to each other rather than trying to become each other.

Accuracy as a feature: what live astronomical computation taught us about trust

Most Panchang tools on the market pull tithi and festival dates from a static, pre-computed table. It's cheaper to build and, most of the time, close enough. We built KaalSetu to compute Tithi, Nakshatra, Yoga and Karana live, using Lahiri ayanamsa and Meeus astronomical algorithms, for the specific date and location a person is asking about — because "close enough" isn't good enough when someone is choosing a muhurat for a wedding or a griha pravesh.

The lesson that's carried into every other platform: in any product where a user is making a real-world decision on the strength of what we show them — a visa timeline on Vuray Travels, a scholarship deadline on Eduledge, a tithi-end time on KaalSetu — precision is a trust signal, not a nice-to-have. We'd rather compute something correctly and slowly than serve a plausible-looking number fast.