A Computer That Watches the School Room: Self-Hosted AI for Homeschool Feedback

From Cedar Park, Texas — August 2026

We homeschool three kids. Like most homeschool parents, our bottleneck was never curriculum. It was feedback. I can pick great material. What I cannot do is sit behind three children at once and see, honestly, how the work actually goes: who is stuck, who is drifting, which problem stopped them cold.

So this summer we built a small referee for the school room. It runs on an Nvidia Jetson Orin Nano Super, a $249 computer the size of a paperback, and after one full week of daily use I want to share what it does, what it caught, and where we are taking the on-device AI next.

What the rig is

Three Logitech webcams, one per desk. A speaker. The Orin. That is nearly the whole list.

The Orin is a real AI computer, not a toy. It has a GPU designed to run neural networks locally, which means every model we use — speech to text, text to speech, vision, a small language model — runs inside our house. Nothing about our kids, their faces, their voices, or their schoolwork ever leaves the room. For a homeschool family, I think that is the most important sentence in this article.

On a typical school morning the system:

Sees. A still frame from each desk every five minutes through the morning block, and continuous video when we run a diagnostic.
Hears. It can speak a math question through the speaker and listen for the answer at the desks, transcribing locally.
Remembers. Everything lands on our home server: contact sheets, videos, session logs, and a running “difficulty ledger” of exactly which skills tripped up which kid.

The philosophy part

The hardware is the easy half. The other half is a daily rhythm borrowed from Japanese craft tradition:

Kata: copy the form. Drill the exact skill.
Poka-yoke: mistake-proof the room. Phones live elsewhere. Books open to the right page before anyone sits down.
Shokunin: the 100% gate. You do not advance on 92%. You repeat the level.
Kaizen: one percent better, one small step, every day.
Hansei: end-of-day reflection. What went well, what broke, one change for tomorrow.

The system runs this loop with the kids; the cameras and logs keep me honest about how it is really going. Together they form a loop any engineer would recognize: observe, orient, decide, act. A homeschool parent’s OODA circle, running every morning before the breakfast dishes are done.

What it caught in week one

This is the part that sold me. We ran a 30-minute math diagnostic with the cameras rolling, then let the system analyze itself: frame extraction, stall detection, and transcription of the kids’ handwritten worksheets.

Our oldest (17). Her linear equations were flawless, eight for eight. But on the factoring sheet she had written, in her own handwriting across the page, “I forgot how to factor.” No drama, no wrong answers to grade. Just an honest self-report that a worksheet full of empty boxes would never have surfaced. She now gets a factoring refresher built around the FOIL steps she half-remembers, instead of a generic review.

The middle (11). Every single answer was correct, 24 out of 24. But the video showed he spent two-thirds of the session motionless, thinking. He computes facts like 12 x 11 from scratch every time. So his problem is not understanding; it is speed. He now does sixty-second fluency sprints where the goal is to beat his own last time, not to beat anyone else.

The youngest (7). Finished everything, passed clean. The camera moments that looked like “stuck” were just a seven-year-old resting between problems. The real finding was for us: thirty minutes is past her attention window, so her next diagnostic gets a shorter block. The data corrected our design, not hers.

None of these came from a test score. All of them came from watching kindly, at scale, and reading what the kids actually wrote.

How the decisions get made

The pipeline is worth describing because the architecture generalizes to any “AI watches a physical process” problem:

video (3 cameras) -> frames -> motion/stall analysis
                  -> two-model vision pass (local VLM + cloud VLM, arbitrated)
                  -> worksheet transcription -> math verified independently
                  -> human review -> difficulty ledger -> next day's drills

The two-model pass matters. A single vision model flagged “phone visible” 23 times on a desk where no phone ever appeared. Running a second model and requiring agreement before acting prevented 23 false accusations against a kid who was quietly doing mental arithmetic the whole time. Any system that grades humans needs an arbitration layer before its verdicts become consequences.

The worksheet transcription is the sleeper feature. Handwriting off a 720p camera, upscaled and read by a vision model, then checked by ordinary code that verifies the math — that is where the diagnostic gold lives. A parent can see which specific problem types broke which specific kid.

The part I almost forgot to mention: the AI colleague

Here is the part that surprises people most. The cameras, the models, the pipeline — none of it runs itself. The analysis I described above was done with an AI agent: a Hermes agent running on my home GPU fleet, powered by DeepSeek V4 Flash served locally, alongside the local Gemma vision model and cloud Gemini when a second opinion is warranted.

I do not mean I click a button and read a report. I mean we work the case together. The agent extracts the frames, runs the motion analysis, and reads the worksheets — then we argue about what it means. It flagged 23 possible phone sightings; I looked at the frames and called them false positives. It transcribed “I forgot how to factor” off a 720p video; we decided together that a self-reported gap is a different thing than a wrong answer. It drafted the refresher worksheets, the ledger entries, the calendar events; I approved, edited, and printed them. The whole diagnostic-to-drills loop closed in about an hour because one of the parties in the conversation never gets tired of re-watching footage.

That is the actual product. Not cameras, not models — a colleague who has seen every morning session and never forgets a detail, running on hardware I own. The homeschool rig is the use case; the pattern is an AI partner with perfect recall of everything that happened in the room, one that helps you teach instead of just automating worksheets.

Where we are taking the on-device AI

Right now the Orin handles capture, audio, and scheduling, and a home server handles the heavier vision passes. The roadmap is to collapse the whole loop onto the Orin itself, because the hardware genuinely supports it:

Parakeet-class STT on-device. Speech-to-text models in the half-billion-parameter range run comfortably on the Orin’s GPU. Kids’ spoken answers transcribed in the room, no network.
A small VLM on-device. Structured frame analysis (present, on-task, phone visible) does not need a frontier model. Quantized 4B-class vision models handle it; the two-model arbitration can stay local too — one model on the Orin, one on the home server.
The quiz brain on-device. A 4B-12B language model is enough to generate, speak, and grade drill questions. We already run this pattern elsewhere in the house.
Neural TTS on-device. The room’s voice, fully offline.

The practical case is threefold: privacy (kids’ faces and voices never leave the house), reliability (the school room works when the internet does not), and speed (no round-trips in the middle of a kid’s attention window). The Orin’s GPU delivers up to 275 TOPS in MAXN mode; a loop of STT + small VLM + small LLM + TTS fits inside that envelope with room to spare. The deep-analysis tier — weekly reviews, long transcripts, narrative video understanding — stays on the home server. The room runs its day; the server runs the week.

What I would tell another family

Cost: roughly $450 to start. Orin Nano Super $249, three used webcams about $25 each, a USB speaker you probably own, and a drive for the recordings. Software is free and open.

Effort: the parts took an afternoon. The software took me a few weekends, but I enjoy this stuff. If you do not, start smaller: one camera, one still per hour, and a folder you actually look at on Friday. The discipline of looking matters more than the pipeline.

The rule that keeps it sane: the computer gathers evidence. The parent makes the calls. Every finding gets confirmed by a human before it becomes an assignment, and the kids know the cameras exist and why. It is not surveillance. It is a lab notebook for learning.

The best feature turned out to be for me. The daily log ends with the kids’ own reflections, and hearing a seven-year-old describe a book as “nice but hard” is better assessment data than any quiz I could write.

We are one week in. The room is quieter, the drills are sharper, and Sunday is now the break day by design, not by exhaustion. That feels like a good trade for a $249 computer and three old webcams.

Gear: Nvidia Jetson Orin Nano Super dev kit, 3x Logitech C920, USB audio. Everything runs locally. Questions welcome.