I Built an AI Learning Companion for My 8-Year-Old. Here’s What Most Kids’ AI Gets Wrong.
Most AI tools for kids treat each conversation like a support ticket — opened, resolved, closed. Spark is different. It remembers, observes, and grows with one child over time.
My son Prabh asked me last week why planes don’t flap their wings like birds do.
It’s a good question. The kind that sounds simple until you try to answer it, and then you’re somewhere between Bernoulli, evolutionary constraints, and the history of flight. He didn’t want a Wikipedia summary. He wanted to think it through. He wanted to argue about it.
Most AI tools would have given him a three-paragraph answer and moved on.
That’s the problem I wanted to solve.
What AI tutoring for kids gets wrong
The market for AI-powered learning tools for children is exploding. Apps that quiz kids, explain concepts, generate worksheets, adapt difficulty levels on the fly. They’re impressive in their own way. They’re also fundamentally transactional.
Every session resets. The AI doesn’t know what Prabh wondered about last Tuesday. It doesn’t know that he lit up when we connected drain swirls to galaxy rotation, or that he goes quiet in a different way when he’s confused-and-curious versus confused-and-defeated. It doesn’t know that his questions have been getting more complex over the last six months — that he’s moved from “what is X” to “how do we even know X is true.”
The tools treat each conversation as a customer support ticket. Opened, resolved, closed.
A real teacher doesn’t do that. A real teacher accumulates a child. They build a living picture of who this kid is, how their mind works, what they cared about three weeks ago that they’ve forgotten they cared about.
None of the AI tools I tested for Prabh do that. So I built one that does.
What I built: Spark, a personal AI learning companion
Spark is a personal AI learning companion designed around one idea. Memory creates meaning.
The conversations Prabh has with Spark aren’t isolated. Every session is observed. Not just stored — analyzed, across four dimensions at once.
Cognitively
Is Prabh’s thinking deepening? There’s a ladder from lookup questions (“what is a black hole?”) to epistemological ones (“how do scientists actually know black holes exist?”). Spark tracks where he is on that ladder and how fast he’s climbing.
Emotionally
Is he engaged or withdrawing? An 8-year-old who types “ok” and goes quiet is telling you something. Curious-confused looks one way. Defeated-confused looks another. Spark notices the difference.
Pedagogically
Is Spark actually teaching, or just answering? There’s a real difference between handing a kid an explanation and asking him to predict something before the explanation arrives. Spark grades its own teaching quality. Not just what Prabh learns, but how much he’s being made to think.
Longitudinally
Six months from now I want to ask a real question. Is Prabh asking better questions than he was in May? That question is meaningless without a baseline captured today. So Spark takes write-once snapshots. Deliberate, timestamped, irreversible anchors that make future comparison possible.
This is the north star. Not engagement metrics. Not quiz scores. The quality of the questions a child asks.
The architecture follows the philosophy
If you build software, this part might interest you. The system is built around the idea that intelligence should be observed, not just served.
After every turn in the conversation, five specialist AI agents run in the background. Each one watches for something specific. One looks for new interests or intellectual threads opening up. Another watches for emotional signals — engagement, frustration, breakthrough moments. A third evaluates the pedagogical quality of Spark’s own response. They run in parallel, each one updating a persistent profile without overwriting what the others have observed.
The profile itself is a living document. Not a log. A structured, growing picture of who Prabh is as a learner. His interests and their trajectories. His cognitive style. His moments of genuine insight. His first-time intellectual moves — the first time he asked a “how do we know” question instead of a “what is” question gets noted, because that’s a milestone.
None of this is exposed to Prabh. He just talks to Spark. But every reply Spark gives is shaped by everything it’s learned about him.
Under the hood: FastAPI and Python on the backend, React and TypeScript on the frontend, SSE streaming for the chat, SQLite for sessions, a crash-safe task queue for the observer agents, and atomic profile updates protected by asyncio locks so the five agents don’t trample each other’s writes. The observers run on a mix of Claude Sonnet and Haiku depending on the depth of analysis each needs. It runs self-hosted. All data stays local.
Why this matters beyond my son
Prabh is 8. He’s curious about black holes, airplane wings, and why the sky is blue when space is black. He doesn’t need a tutor. He needs a mind that stays. One that remembers what he cared about, notices when he’s growing, and meets him where he is.
The technology to build this has existed for a year. What’s been missing is the conviction that children deserve AI that treats them as evolving, complex humans rather than query senders to be answered and dismissed.
Every parent knows that their child’s questions change over time. They get deeper. More abstract. More self-aware. That change is the whole point of education.
The question is whether the tools we build for children can actually see it.
Spark is my attempt to make sure at least one tool can.
Frequently asked questions
What is an AI learning companion, and how is it different from an AI tutor?
An AI tutor typically answers questions or walks a child through a curriculum. An AI learning companion goes further. It builds a persistent understanding of who the child is as a thinker — their interests, their cognitive habits, the questions they’ve asked over weeks and months — and shapes future conversations using that context. The difference is memory and observation.
How does Spark remember conversations across sessions?
After every turn, five specialist AI agents observe the conversation in parallel — each one looking for a different kind of signal (new interests, emotional state, pedagogical quality, milestone moments, longitudinal change). Their findings update a structured profile that grows over time. The next conversation is informed by everything in that profile.
Is Spark safe for children’s data?
Spark runs on a self-hosted instance. All conversation data and profile data stay local to the deployment. No third-party analytics, no advertising trackers, no data sold to anyone. It was built for one child, and the privacy model reflects that.
Why use multiple AI agents instead of one?
One agent trying to track everything at once produces shallow, hedged observations. Five specialist agents each watching for one thing produce sharper signals. They run in parallel after each turn, so the system stays fast, and they write to the profile atomically so their observations don’t conflict.
What’s the tech stack behind Spark?
FastAPI and Python on the backend, React and TypeScript with Vite on the frontend, SSE for streaming chat responses, SQLite for session storage, a crash-safe task queue for the background observer agents, and asyncio locks for atomic profile updates. The observer agents use a mix of Claude Sonnet and Haiku.
Could a system like this work in schools?
The architecture would scale to multiple learners, but the design philosophy assumes a sustained relationship between one AI and one child over time. A classroom deployment would need careful thought about how teachers see (or don’t see) the profile, what’s shared with parents, and how to keep the observation focused on growth rather than evaluation.
Is Spark available to use?
Not currently. Spark is a personal project, built and deployed for one learner. The technical write-up here is meant to share the design ideas, not pitch a product.
Leave a Reply