Independent Builder Project

Introducing INKBOT

An experiment in “intelligence architecture,” human intent, and making AI output more inspectable before it becomes trusted.

In One Sentence
INKBOT is a browser-based prototype for translating ordinary human language into structured, reviewable AI handoffs before an AI is asked to generate, analyze, or act.

Thanks for the context and examples, @forum-helper. I wanted to share an independent prototype I have been developing called INKBOT.

It began with a specific problem I repeatedly encountered as a builder: the friction between what a person naturally means and what a multimodal AI system infers. Models are excellent at producing raw outputs, but getting to a clear, accurate, human-meaningful interpretation often becomes a separate alignment problem.

Instead of asking users to learn prompt engineering, INKBOT acts as an intermediate intelligence architecture layer that structures the human meaning between human intent and AI output. Its working loop is:

Core Intent Loop
DESCRIBE → MAKE IT VISIBLE → RECOGNIZE → CORRECT → REFINE

Figure 1 — INKBOT Intelligence Architecture. A visual representation of the prototype’s central loop and supporting structures: human intent, visible output, inspection, correction, concept memory, references, provenance, validation, and revision state.

https://ko-fi.com/thomascoates/shop

I have put together an evaluation package containing complete runnable local HTML files for both core versions of the prototype:

Entry-Level Experience

INKBOT Lite 71

The entry-level abstraction. It takes about five minutes to test: users describe concepts in ordinary language, receive a usable visual handoff, and can inspect whether the first picture reflects what they actually meant.

System / Research Experience

The Intelligence Architecture Edition

This sits beneath the simple interface. It explores what happens when a user wraps a model’s generation loop in explicit programmatic structures such as concept memory, references, provenance validation, local database retrieval, revision histories, and version histories.

What I am testing

• Can an AI help a person make an idea visible before committing to a large output?
• Can users inspect and correct model interpretation instead of trying to repair the final result after the fact?
• Can provenance, references, validation, and revision state make multi-step AI work easier to trust and audit?
• Can a simple visual workflow hide technical complexity without hiding the meaning being preserved?

One of the practical case studies included in the package involves a complex shore-shell mapping workflow: a field technician, a survey wheel, sequential photography, GPS mapping coordinates, and overlapping frames. INKBOT successfully translated that multi-step workflow into a clean, unified system diagram that can be inspected visually before execution.

The visual below is not intended as a claim that the implementation is complete or production-ready. It is a concrete, inspectable artifact of the kind of human-to-AI translation problem I am exploring.

Suggested Additional Visuals for the Evaluation Package

I am expanding the package with additional visual evidence so reviewers can inspect the workflow from multiple angles rather than relying on product claims alone.

1. First-picture handoff example. An ordinary-language idea, the structured handoff, and the resulting first generated image shown side by side.
2. Human correction path. A before-and-after example showing how a user identifies a wrong relationship, object role, or spatial arrangement and corrects it without rewriting the entire idea.
3. Provenance and revision view. A visual audit trail showing what came from the human, what was inferred by the AI, and what was subsequently approved or changed.
4. Beach Shell Mapping storyboard. The full physical-to-digital workflow that demonstrates concept translation across movement, imaging, location data, selection, and retrieval.
Public Prototype and Project Links
INKBOT Lite 71
https://erniewood.neocities.org/INKBOT/InkBotLite71
INKBOT 3 — Intelligence Architecture Edition
https://erniewood.neocities.org/INKBOT/INKBOT3IntelligenceArchitectureEdition
Project / Support Page
Replace with your current Ko-fi or public project URL before posting.

I am asking the community to critically inspect the design rather than simply accept the framing. I would welcome feedback from people working on multimodal systems, instruction following, model behavior, evaluation, human–AI interaction, safety, or agentic workflows.

In particular: where does this duplicate existing work, where does the architecture become genuinely useful, what important limitations am I missing, and where would experienced builders simplify or challenge the design?

Questions I would value feedback on

• What existing tools, papers, products, or frameworks most closely overlap with this approach?
• Is the separation of human-approved meaning from AI inference useful in practice, or unnecessarily heavy?
• Which parts are best treated as interface design, which as application logic, and which as model behavior?
• What would make this prototype a stronger, more honest evaluation artifact?

Thanks for taking the time to look at an independent community project.