Constellate

AI makes it effortless to stop thinking, just as judgement becomes the only job left. Constellate is the other future: an AI that teaches the dots but refuses to think for you. Your sky grows.

Comment

Inspiration

The founding moment was three unrelated dots colliding: reading Steven Strogatz taking calculus apart until it felt inevitable, a long conversation with an AI about why the oldest teaching method might finally be scalable, and hearing Terence Tao reach for a Simpsons line, "I've tried nothing and I'm all out of ideas," which I cannot stop hearing as the entrepreneurial tag line and education reform in one sentence.

Three dots, one line drawn between them: that is the whole theory of what a mind is for, and it is also, literally, how this product happened. The line I wrote at the end of that founding conversation is still the test: "I feel genuinely emotional now, energised and graduated. If we can recreate this for even one individual we have succeeded." This XPRIZE deadline lit the fuse. Then the harder question underneath: Can AI teach, judge, author and run the whole business too?

What it does

Constellate teaches concepts as real guided, interactive and adaptive conversations. The guide doesn't lecture. It asks, builds from your own world, and closes the session only when you have demonstrated you can wield the idea in your own words. Each earned idea lights a star on a personal map of understanding allowing you to see how your understanding has connected, built and evolved.

Behind the curtain, AI judges every session on both sides of the table, guide and learner, authors new concepts where learner demand pushes, and runs the business unattended every morning: collecting judge verdicts, reading learners, drafting the day's improvement insight and marketing, and emailing the one human whose job is to approve, edit, or kill. Everything is live in production generating initial revenue.

How I built it

With AI, all the way down. AI as a thought partner. AI as a guide. AI as a coding agent and an AI runs the product. Last count we had 7 different models running in Constellate, each with a specific role and function, with plans to add many more.

Challenges I ran into

The hardest challenges were making an AI relatable as a guide, and honest as a learner when we generated synthetic practice sessions. Early generations of guides would capitulate: ask a beautiful question, get silence, then answer it themselves, then graduate a learner who had demonstrated nothing. Testers split on tone: some found the guide dismissive where others praised the connection, and both were reacting to the same prompt. And the first synthetic learners were nothing like real humans, so we ran loop after loop of scored sessions to learn where synthetic data could be trusted and where only real learners count. AI is jagged; smoothing it is the job.

Accomplishments I'm proud of

Delivery. Many late nights and long weekends to ship this, and I am sure anyone that builds in their 'spare' time knows it is not an easy task; all of us have our fingers crossed that maybe one day it becomes the day job.

What I learned

Generation can be cheap and judgement is expensive, so we engineered judgement into every layer instead of treating it as vibes. Honesty is a strategy, not a constraint: an AI that admits what it cannot verify, refuses to grade thin evidence, and reports weaknesses is more trustworthy than one that performs certainty.

Measurement is the real product: the moment sessions are scored, the system can adapt to learners and improve itself faster than we can improve it by hand. And distribution, not capability, is the gap between a working engine and a business. Generating revenue earlier here was key and a missed opportunity but 90 days to build, ship and generate was a high bar.

What's next for Constellate

Now the engine is built, any concept and learner map is possible, making the differentiator the ability to create, map and process more knowledge, while the AI continually adapts and evolves.

Future avenues include educators buying evidence-quoted portraits of how students think, organisations mapping where understanding really sits in their teams. Entrepreneurs and learners wanting to create their own pathways, Constellate agents supporting consistency, and of course trying to keep riding the incredible improvements that AI is making on a daily basis to be on the right side of the wave.

Finally, with model improvement driving costs down, Jevons economics should let us put frontier models in every session and generate voice and video on the fly. An exciting future!

Share this project:

Updates