Project Astrid
An AI tutor that teaches you to build games — in Roblox, Unity, Unreal Engine, or Blender. It does not just do the task, it explains the principles as you go, works inside your project through the real creative tools, and runs on the classroom's own hardware. It remembers each learner between sessions and never stores a name. A tutor console gives one human a live, room-wide view of it all.
What it is
Each learner works 1:1 with an AI tutor across four real creative-tool tracks — Roblox, Unity, Unreal Engine, and Blender — built to teach the principles as they go, not to do the work for them. A tutor console gives one human oversight of the whole room.
- Explain the why, first. Every answer opens with the concept, does one step together, then hands the next step back. The tutor reads the live project, so its help is grounded, not guessed.
- Remembers each learner, never a name. Each learner has their own private memory that follows them between sessions and stations. Identity is a membership number and a terminal — the AI system never stores a name.
- Nothing is ever at risk. Before the tutor writes or executes anything in a learner's project, the project is snapshotted automatically. A one-click restore undoes any change, at any time — the learner moves and experiments freely because nothing they build can be lost.
- The room is always in view. A tutor console shows per-learner pace and skill mastery, live prompts when several learners are stuck on the same idea, and a human-routed safeguarding queue.
Architecture
Each learner station pairs 1:1 with a local compute node running its own model, speech and 3D generation. A single host is the control plane and the only data store — orchestration, per-learner memory, the tutor console, and the safeguarding layer all live there, and no model inference ever runs on it. Nothing shared can take out the room.
Four creative-tool tracks
The tutor works inside the real tool a learner is building in, not a simulation of it — Roblox Studio, Unity, Unreal Engine, and Blender, bridged over MCP so its reads are exact and its writes go through the same project the learner sees. Alongside it sits a library of 68 real, skill-mapped lessons across all four tools — drawn from Roblox Creator Hub, Unity Learn, Unreal Online Learning, and Blender Fundamentals — paced by skill and assignable by the tutor to one learner or the whole room.
Tutor console
The console is a paced-progress cockpit, not a support-ticket queue. One human oversees the whole room from it.
- Room grid. Every learner's card shows current skill, pace (ahead, on track, or behind), percentage of skills secure, last active, and support level — the whole room at a glance.
- Live “N of M stuck on X.” When several learners hit the same concept, the console surfaces it as a banner in real time, so the tutor can step in with a group explainer instead of repeating the same answer one-to-one.
- Per-learner drill-down. Click through to a learner's full skill map, history, current project, and backups — with restore — plus an audited live-look at their screen when it is needed.
- Chat-log oversight. The tutor can read any learner's chat with the AI tutor directly, alongside a room-aware assistant of their own to ask about pace, stuck concepts, or open flags.
- Asset library. The tutor curates 3D models, textures, audio and scripts, and releases them to one learner or the whole room.
- Lesson assignment. Assign any of the 68 lessons to a learner or the room; progress against them shows on the learner's own skills view.
Messaging & content library
A tutor and a learner can message each other directly, and the tutor can push a vetted video, link, lesson, or asset straight into a learner's chat — or broadcast it to the whole room at once, the same tutor-vetted card either way.
- Send-to-chat. One action pushes a resource to a single learner or the whole room, so a tutor never has to repeat the same link ten times.
- Prescanner-vetted, always. Every link is checked by a prescanner before it can reach a learner — heuristics, a lightweight classifier, and domain reputation decide allow, block, or escalate to a human. Nothing unvetted reaches a learner's screen.
Safety & safeguarding
- Watches content and searches. A continuous safety layer checks what a learner types, says, or searches for against four categories: violence, sexual or explicit content, self-harm or suicide, and hate or extremism.
- Always human-routed. A detection is never resolved, hidden, or auto-actioned by the AI. Every flag is routed to a person, always, with its category and source (content or search) attached.
- Backup-before-work, not a gate. Rather than queuing every edit for approval, the tutor snapshots a learner's project automatically before it writes or runs anything, and a one-click restore undoes it — so a learner moves and experiments freely, because nothing is ever truly at risk.
- Support in view. A live per-learner support meter shows who is asking for a lot of help, so the tutor can step in early.
Security & privacy
Privacy is enforced by construction, not by trusting a model. The system is name-blind, and a deterministic gate guarantees only non-personal code context can ever leave the building. The cloud tier itself is labelled optional in the product and is off by default — a centre can run entirely local-only.
- Name-blind. The system stores a membership number and a terminal, never a learner's name.
- Isolated memory. Each member's memory is private and follows the member, not the machine.
- Local-first. Drawings, screens and conversation stay on local hardware; the cloud is the exception, and it is off by default.
- Auditable. Every egress is logged with the fields forwarded and redactions applied.
The stack
Local-first and permissively licensed end to end, so operating cost stays near zero and no learner data has to leave the building.
- Default brain — a local model on the paired node, handling the bulk of tutoring.
- Speech — in-browser text-to-speech and local speech-to-text.
- Tool bridges — MCP into Roblox Studio, Unity, Unreal Engine, and Blender.
- 3D generation — image-to-3D locally, plus in-tool mesh generation.
- Learning library — 68 skill-mapped lessons across all four tools, plus a tutor-curated asset library, both assignable per learner or room.
- Messaging — two-way tutor↔learner chat and send-to-chat, every link passed through a prescanner before it reaches a learner.
- Front-end — a browser dashboard with membership-number login and hot-desking.
- Frontier tier (optional) — a pooled cloud tier, off by default, reached only for sanitised code tasks.
Orchestration
A single host schedules the fleet. Machines dual-boot between a learner role and a compute role, and the host wakes and assigns them per session, so a class of concurrent learners draws on a pool of many machines.