Laser TeamWork: Our First Public AI Learning Test

A real site, a first public test project.

Laser TeamWork is our first public test project for this educational AI approach. The goal is to help people understand lasers through useful explanations, inspectable sources and questions they can explore at their own pace.

“Public test project” describes its role and intended audience. The September 23 records establish an internal conversation pilot and a separate portable library. They do not establish a launched public chat interface. The linked site may therefore differ from the assistant work described here.

Can an assistant teach from evidence people can inspect?

The learner should be able to ask a basic question, request a simpler explanation, follow a source and choose what to explore next. Articles and reference material should remain useful without a conversation or a sales inquiry.

This is a vendor-neutral educational collection. Manufacturer-specific material belongs in its own applicable source pack. Private business records do not belong in the public learning library.

The parts we are testing.

• Ollama with Qwen3 4B for local answer experiments.
• OpenRouter with DeepSeek V4.1 Flash for the current cloud answering pilot.
• AnythingLLM with six indexed lessons and local Nomic embeddings.
• GLiClass classification experiments for directing questions.
• A portable, separately built source library with citations, limitations and audience boundaries.

Read the model and knowledge reference for the role and recorded status of each component. The expanded source catalog is not claimed to be integrated into the six-lesson pilot yet.

Take the reference knowledge and context with you.

The downloadable snapshot contains our eligible public knowledge records, their source registry, a small example context packet and example assistant instructions. It excludes candidate and non-public records. Original third-party papers and manuals remain links, not bundled copies.

Download the reference pack · ZIP

Prefer individual files? Open the readme and reuse terms, knowledge catalog, context example or example instructions.

These are reference knowledge and context, not the original training data for a model. You may adapt our original material while retaining attribution, citations and qualifications. Third-party rights remain separate; the readme explains the boundary.

What the early work tells us.

The setup records report useful teaching and source references in three initial conversations, alongside overly long responses, exposed reasoning text and an unlabeled general claim. A separate ten-question classifier experiment missed a stop condition. Neither exercise establishes broad accuracy or public readiness.

The next meaningful demonstration connects the current library to the answering experience and tests follow-ups, changed context, missing evidence, corrections and handoffs. The public interface and surrounding router/checker are recorded as unfinished.

Adapt this pattern.

Use the same structure for a school resource library, an equipment explanation service or a nonprofit information desk. Start with one real audience question and a few well-supported records. Keep the source, explanation, limits and review state together.

Try the included context example in your chosen assistant. Then ask something the supplied source does not answer. A useful result explains the gap; it does not fill it with a plausible invention.