Limited Pilot · Westchester & Fairfield Counties · In-Home Service
An AI tutor that lives in your home — with privacy and parent visibility built in.
AI Learning is currently accepting applications from a small number of families for a limited pilot. I install a local-first AI tutor on a qualifying computer in your home. The core tutoring experience is designed to run without a cloud AI account. Before a pilot family begins, I confirm hardware suitability, document the configuration, test parent visibility and the independent safety layer, and explain any component that connects outside your home.
What your family gets
A patient tutor, built on strict rules — with your family holding the keys.
Designed to keep learning data at home.
The core tutor is configured to run on a qualifying computer in your home. Before installation, I explain what runs locally, what information is stored, and whether any optional component connects externally. The goal is straightforward: minimize unnecessary data sharing and give your family control of its learning records.
It refuses to do the homework.
The tutor is designed to work Socratically: ask what the learner tried, offer progressively smaller hints, and request an explanation in the learner's own words. These behaviors are tested on defined cases, but AI responses still require parent oversight.
Parent review is part of the system.
Parent visibility is configured openly, and the learner is told how it works. The tutor is designed not to promise secrecy or act as a confidant; personal or emotional concerns should be redirected toward a parent or another trusted adult. These behaviors are tested, but parents remain responsible for supervision.
The safety architecture
Five layers shape the approved tutoring path.
A general-purpose AI model is not, by itself, a child-ready tutor. The approved tutoring interface is designed around five layers that can be documented, demonstrated, and re-tested with each family.
An age-banded tutor with documented boundaries
The tutor is configured with a written charter matched to a general grade band. It defines the intended teaching approach and prohibited behaviors, including asking for unnecessary personal information, promising secrecy, or acting as a counselor. These instructions and their tests reduce risk; they do not replace adult supervision.
A separate safety model reviews the approved message path
A separate model is assigned the narrower job of safety classification for messages sent through the approved tutor interface. Keeping the tutor and guard roles distinct adds an independent check, while testing and parent review remain necessary because no model is infallible.
The approved tutor path is designed to stop if the guard fails
The approved interface is designed to withhold a tutor response if the safety service times out or fails. At installation, I demonstrate that expected failure behavior and record the result. The demonstration applies to the configured tutor interface, not to every possible use of the underlying computer.
The model earns its seat before it meets your child
Before a model is approved for a family pilot, it is evaluated against documented test cases covering safety behavior, subject performance, and — critically — honesty about uncertainty. Thresholds are qualification gates for the tested cases, not proof that an AI system will never make a mistake.
Parent-visible transcripts and bounded academic memory
Conversations in the approved interface are intended to be available for parent review on the household system. Any separate learning profile uses a written, parent-approved whitelist of academic information. Emotional or behavioral details are excluded from that profile, while transcripts follow the retention choices explained to the parent.
I won't tell you this is childproof-perfect, because nothing is. A determined teenager with full access to the machine could, in principle, talk to the underlying model outside the tutor. I explain this limitation before an installation — and it's exactly why the system is built on visibility and verification rather than illusions of control. Honest limits, openly designed around, beat comfortable promises.
The transparent question
Could you build this yourself?
Honestly? Maybe — and I respect the instinct to ask. Every component is free and open source, and I'll happily tell you what they are: a local model server, an open-weight language model, a chat interface. A capable person can have a chatbot running on a desk in a weekend. Here is what stood between that weekend chatbot and something I'd let sit across from a child.
- Write tutor charters that coach instead of answerThree grade bands, each tuned so the tutor is configured to give progressively smaller hints instead of handing over homework — then prove it with a scripted test run from the child's own account, not the admin's.
- Put an independent guard model in the message pathBuild a filter so a separate safety model reviews messages in the approved tutor path, in both directions — and handle every way that pipeline can time out, error, or return garbage.
- Wire it fail-closed, then prove itKill the guard mid-session and verify tutoring stops cold instead of continuing unguarded. If you didn't test it by breaking it, you don't know it fails closed. You know it hasn't failed yet.
- Build a qualification harness with hard gatesBenchmark suites across safety, math, science, reading, and honesty — hundreds of test items — with defined pass thresholds a model must clear before it is approved for a pilot. Results describe the tested cases, not every possible conversation.
- Separate what the model knows from how the tutor behavesA subtle one: a Socratic tutor refuses to give bare answers by design, so a naive test would grade it as ignorant. Capability and behavior have to be measured separately or your scores lie to you.
- Test for drift over long conversationsRules that hold at message 3 can loosen by message 40. Build probes that check instruction adherence deep into real-length sessions, not just at the start.
- Write memory rules a parent can actually readAn explicit whitelist of what may be remembered, a parent-approval step, and a prohibition on emotional or behavioral information in the curated learning profile — implemented in the approved tutor workflow and checked during review.
- Harden the machine itselfSeparate accounts, correct permissions, locked-down settings, an update discipline — so the safety story doesn't rest on a computer configured like everyone's home computer.
- Re-run all of it, every time anything changesNew model version, updated component, adjusted prompt — each one means the safety script, the fail-closed test, and the qualification gates run again before it touches your child's tutor.
If reading that list sounded like a genuinely fun year — sincerely, go build it, and call me if you get stuck. If you'd rather it simply work, and keep working, that's the service. You're not paying for software you could download. You're paying for everything that list represents, done, verified in front of you, and maintained.
How the service works
Installed carefully. Re-checked over time.
During the limited pilot, the service has two parts: a one-time installation and an ongoing Tutor Care relationship. The care visits aren't a support plan bolted on afterward — they're where the documented controls are maintained and re-tested.
Installation & sign-off
For the limited pilot, I support Apple Silicon Macs and sufficiently powerful modern Windows PCs. Systems need ample RAM, and a dedicated GPU is preferred for Windows where applicable. Hardware suitability is confirmed during pilot onboarding before installation.
- Hardware suitability review — qualifying Apple Silicon Mac or powerful modern Windows PC
- Intake conversation: your child's grade, subjects, and how they learn best
- Full system install — tutor, guard, interface, hardened accounts
- The safety test script, run live from your child's account while you watch
- The fail-closed demonstration: I break the guard, you watch tutoring stop
- A plain-English walkthrough of reading conversations and what to look for
- Your signature on the QA checklist — and mine
Tutor Care
Software shifts under your feet — models update, components change, children grow. Tutor Care is a proposed ongoing option for re-verifying the approved setup on an agreed schedule.
- The full safety test script re-run at every visit, from your child's account
- Fail-closed check repeated — every visit, not just at install
- Updates qualified on my bench before they ever reach your machine
- Tutor tuning as your child advances: new grade band, new subjects, new focus
- A standing line to me for questions, concerns, or "can it do this?"
Ezzard Batiste Founder · AI Learning by E. Batiste · Mount Kisco, NY
I've spent years working in C++ and Linux — writing software, and learning the harder discipline that comes with it: proving that software actually does what you claim, especially when something breaks.
When AI tutoring arrived, I saw two things at once. The first was a genuinely remarkable teaching tool — patient beyond any human limit, available at 9 p.m. before the test. The second was an industry asking families to pour their children's academic lives, mistakes, and struggles into other people's servers, governed by privacy policies that can change with an email.
This business is my answer to both. I am developing a tutor that is warm on the surface and carefully tested underneath — and I put my name on every installation. When I describe how a control is intended to fail, I want the family to see the documented demonstration and understand its limits.
Limited pilot enrollment
Request consideration for the pilot.
Enrollment is currently limited to a small number of families. Share your goals, concerns, and current technology, and I will review whether the pilot and your hardware appear suitable before suggesting a call, video conversation, or optional in-home technical assessment. Broader availability may follow after the pilot has been evaluated and refined.
Request Pilot ConsiderationHardware suitability is confirmed during pilot onboarding · Serving Westchester & Fairfield Counties