Charter School Technology Intelligence

Market intelligence for a New Delhi-based software engineering firm

A software firm wanted to build for the U.S. charter-school market. Before they committed, they needed to know a deceptively simple thing: which technology do these schools actually run — the student and learning platforms behind the classroom — and where is the real opportunity? The answer they'd been working from was the number everyone in the market quotes. It was wrong.


The problem

The obvious signal made one platform look dominant. But it was measuring the wrong thing — presence of a popular email and identity suite, not the platform a school teaches on. Worse, most of the real answer lives behind school login walls that no public scan can reach. So the market had settled on a figure that was high in one sense, near-zero in another, and defensible in neither.

The client didn't need a bigger scrape. They needed an answer they could take into a decision and stand behind — with its uncertainty stated honestly, not buried.

What I did

I treated it as two questions, not one: what can we prove, and what must we estimate — and I kept them rigorously separate.

  • A provenance-first system. Every claim is graded by the strength of its evidence — directly verified, inferred, or merely suggested — and the share we simply don't know is reported as a number, not hidden. Nothing is ever promoted from a weak signal to a stated fact.
  • AI agents at national scale. Rather than do the detection and outreach by hand, I designed the workflow and pointed a set of AI agents at it — reading public evidence across the full national universe of schools and resolving the right people to confirm with, where confirmation was possible.
  • A calibration study, not a guess. I hand-verified a representative random sample to measure how often the public signal was actually right, then used that to convert a noisy indicator into an honest range with explicit confidence — and checked the result against an independent third-party dataset I hadn't tuned to. The two agreed.

The result

One defensible answer the client could stand behind — a bias-corrected range that an outside source independently landed inside — and, just as important, a clean separation of two claims that are easy to conflate: what's widely used versus what's the system of record. The two tell very different competitive stories.

It shipped as a live, self-updating report rather than a slide that goes stale the week after delivery.

The impact

It replaced a market-sizing number that was quietly wrong — in both directions — with one that could carry a real decision. It reframed the competitive picture: the "winner" everyone named turned out to be a mirage, and the genuine contest was somewhere else entirely. And it left behind reusable capability — a detection-and-verification method the client can point at the next question, instead of a one-off report.

What it taught me

The hard part was never the analysis. It was deciding which question was worth answering, and how much certainty it actually needed. Precision where it matters, an honest "we don't know" everywhere else — that beats false confidence every time, and it's what makes a number safe to bet on.

And at national scale, the job wasn't to do the work by hand. It was to design the system, point the agents, and own the judgment in between — the reframes, the calls on what to trust, the decision about what "good enough to act on" means. That part doesn't automate.