Five coordinated components. One persistent brain. A custom model that lives in your browser. Here is what happens between “tell us what you're making” and your published draft.
Most AI products are a chat box in front of someone else's model. RetroHub is the NeuroNetwork — five named layers, each doing the job it's best at, all working together.
Every chat turn, every image request, every voice call flows through NeuroCORE. It enforces the workflow stage (Create / Ship / Refine), picks the right provider, applies safety filters, and emits the brand label you see in the chat header. One chokepoint, every decision auditable.
A typed, versioned knowledge graph. Voice nodes, style nodes, character bibles, project context, library items — all of it persistent, scoped to you, and pinned to specific versions so a system update never silently rewrites your in-flight work.
Quill (our 135M custom model) runs in your browser via WebGPU. Cloud frontier models (Grok, Gemini, Claude) handle Ship & Refine. NeuroEngine is the substrate that abstracts both behind one streaming interface.
Each skill — Children's Book, Cold Email, SEO Cluster, Resume — is a 1000-1500 word playbook stored as a NeuroCortex node. At every turn, NeuroSkills selects the relevant playbook plus your in-scope voice and project context, then assembles the system prompt.
The Production Shell — chat-first, with collapsible Brief / Preview / Insights / Personality panels that fade in and out around the conversation. The Brain Window lets you edit NeuroCortex directly. The slash rail surfaces every skill, app, and shortcut.
Five turns. Five components. One coordinated flow.
A blog post about open-source. A children's book chapter where Maple meets the river. A cold outreach email. You type it into the chat input. NeuroUI captures the intent and hands the turn off.
The dispatcher reads your tier, your workflow stage (Create / Ship / Refine), and the task hint. It decides whether this turn goes to the local model, a cloud model, or an image pipeline — and which one. The decision is logged and auditable.
The relevant skill playbook loads (Children's Book, Cold Email, whatever). NeuroSkills walks your NeuroCortex, pulls your in-scope voice, project, and character nodes, and assembles a curated system prompt — version-pinned and deduplicated.
Quill — our custom 135M-parameter model — runs in your browser via WebGPU. Tokens stream into the chat panel at conversational pace. No round trip, no charge, no rate limit. The same NeuroEngine routes Ship turns to frontier cloud models when you commit.
The conversation feeds back into your knowledge graph. Tone corrections become candidate STYLE updates. Named characters become BOOK_CHARACTER nodes. Approved outputs become LIBRARY_ITEMs. Every session compounds — without you doing anything explicit to teach the system.
Other AI tools train you to expect the first response to be the best response. Type a prompt, get a draft, ask for refinement, done. The model is the variable.
RetroHub flips that. The model is a constant. What changes from turn to turn is your working set— the slice of NeuroCortex that NeuroSkills loads as context. Early turns have a less-converged working set than later turns. By the third or fourth draft, the same model is producing dramatically better output, because it finally knows what you actually want.
This is why drafting is free. You're not asking the model to try harder — you're building the context that makes it good. The first draft is the system meeting you. By the third draft, it's working for you.
When you click Ship, NeuroCORE hands that converged context off to a frontier model — and the difference between "chat with GPT" and "chat with the right context loaded" is the entire product.
Every session adds to your NeuroCortex. Every correction tunes your voice. Every project becomes a reference point for the next.
Your NeuroCortex survives every session. The character you wrote yesterday is still there next month.
Your VOICE node grows with you. Sessions pin to the version they were created against — no silent drift.
Pick up the children's book you started last summer. Same characters. Same tone. Loaded automatically.
We don't make money on you exploring. We make money on you shipping.
RetroHubAI-Quill-135M is our custom-trained draft model — the local-first half of NeuroEngine. A LoRA fine-tune of HuggingFace's SmolLM2-135M base, distilled on responses from grok-3-mini, built specifically for the conversational, production-studio voice this product needs.
It runs entirely in your browser via WebGPU + MLC WebLLM — small enough for consumer hardware, fast enough to feel conversational, and tuned for the specific job of drafting alongside NeuroCortex. Every drafting turn at every tier uses it. Free, unlimited, no round trip.
Off-the-shelf 135M models either feel robotic (Instruct variants) or generic (raw base). Quill is neither. It's trained for this product, for this user, for this moment in your conversation.
Drafting is free. NeuroCortex is unlimited. You only pay when you're ready to commit.
