📖 Brand — Adapta¶

Kind: Reference. This document defines what the brand is — name, positioning, voice, color, type, and logo. It carries no open tasks. The work to apply this system to the product lives in the roadmap: TODO.md §B — Brand rollout.

Single source of truth. When a color, font, or string here disagrees with the running product, the product is wrong — fix the product, not this doc. The design tokens in adapta/console/src/tokens.css and the palette in mkdocs.yml are generated from this page.

0. Project status¶

Adapta is a one-person project with a significant portion of its implementation written using AI-assisted code generation (Claude). It is at prototype maturity: the pieces are all wired up — RAG, LoRA training, eval gate, multi-tenant serving, image-understanding fine-tunes — but the GPU training path has limited CI coverage, and the system has not been validated at real team scale or independently audited.

If you're considering deploying this with sensitive data, read the relevant code paths yourself. AI-generated code is a reasonable starting point, not a correctness guarantee. The auth, error handling, and eval gate paths in particular deserve independent review.

This is stated here because the product's core promise is data privacy and eval-gated serving. The gap between what the system does mechanically and what has been independently tested should be clear upfront.

1. Brand in one line¶

Adapta — private model customization, on your own hardware.

The name comes from adapt (Spanish adaptar): the product adapts a general base model to your data — a LoRA adapter changes how it behaves, retrieval adapts what it knows, and the two compose on one endpoint. Short, pronounceable, and literal about what the system does.

Positioning statement¶

For teams that can't send training data to a cloud API: run this on your own hardware. Upload documents for cited retrieval (RAG). Train a LoRA adapter on your data to change how the model responds. Combine both on the same endpoint. An adapter that doesn't clear the held-out eval gate is blocked from serving — mechanically, not by policy. No telemetry; embeddings and vectors stay local.

The specific claim is narrow: the composition. RAG tooling is a crowded space (Dify, RAGFlow, AnythingLLM, many others with large teams and years of production hardening). Fine-tuning tooling is also crowded (Unsloth, LLaMA-Factory, Axolotl). The combination — retrieval + behavior training + eval gate, behind one OpenAI-compatible endpoint, self-hosted — is less crowded. That is the differentiator, not the individual pieces.

What this is not: not a workflow/agent builder, not a deep-document parser, not a chat UI. It is the model-customization-and-serving layer. Don't enter comparisons the project wasn't designed to win.

Taglines¶

Use	Line
Primary (hero, console login, README)	Your model. Your data. Your servers.
Product idea (sub-hero)	Adapt a base model to your data.
RAG / Knowledge mode	Give it your knowledge.
Fine-tune / Behavior mode	Change how it behaves.
Trust / safety (eval gate)	Nothing unverified ever serves.

The two mode lines are locked product framing, not marketing decoration: the UI never calls RAG "training." See PRODUCT_DEFINITION.md and the voice rules below.

2. Voice & tone¶

The product already has a voice: developer-direct, precise, assumption-light, never hyped. The brand keeps it. We are infrastructure, not a growth-hacked SaaS.

Principles

Verbs, not adverbs. We upload, index, train, compose, serve, evaluate. We never unlock, supercharge, harness, or empower.
State the mechanism. "Held-out eval, response-only loss" beats "trusted AI." Naming the machinery is the reassurance for this audience.
Privacy is stated flatly, once, and not oversold. "Your data never leaves your infrastructure." No fear-mongering, no badges of compliance we don't have.
Two modes, two questions — never merged. Specializing a model is always framed as a choice between Give it knowledge (RAG) and Change how it behaves (fine-tuning). Do not call RAG "training," do not call fine-tuning "uploading knowledge." This is a product rule.
Errors are honest and typed. User-facing copy mirrors the DomainError discipline: a safe, plain message — never a stack trace, never false comfort.
State project maturity plainly. This is one-person, AI-assisted work at prototype maturity. Copy does not claim production-readiness the codebase hasn't earned. Where something is early-stage or untested at scale, name it — the audience will find out anyway.

Voice quick-test — if a sentence would feel at home in a YC pitch deck, rewrite it.

✅ On-brand	🚫 Off-brand
"Index your documents. Answers come back cited."	"Unlock the power of your knowledge base!"
"Fine-tuning runs on your GPU. It fails fast if there isn't one."	"Seamlessly supercharge your AI workflows."
"An adapter that doesn't clear the eval gate can't serve."	"Enterprise-grade, battle-tested reliability."
"Early-stage. Read the auth and eval-gate code before deploying with sensitive data."	"Production-ready, hardened, trusted by teams."

3. Color¶

A dark-first system (the console and docs both default to dark). The palette unifies the two surfaces that disagree today — the console blue (#5b8cff) and the docs deep-purple — onto one Indigo brand hue, and adds the brand's most distinctive idea: the two modes are color-coded.

3.1 Core¶

Token	Hex	Role
`--brand` (a.k.a. `--accent`)	`#6366F1`	Indigo — the one brand color. Buttons, links, focus, active nav, progress.
`--brand-weak`	`#1B1F3A`	Indigo tint for selected/active backgrounds on dark.
`--brand-ink`	`#FFFFFF`	Text on a solid `--brand` fill.

3.2 The two modes (the signature of the system)¶

The product's whole shape is knowledge vs behavior. Encode it in color so a glance at a project card tells you what kind of project it is.

Token	Hex	Mode	Meaning
`--knowledge`	`#2DD4BF`	Knowledge (RAG)	Teal = grounded, factual, cited.
`--knowledge-weak`	`#0E2E2A`		tint background
`--behavior`	`#A855F7`	Behavior (fine-tune)	Purple = learned style/skill.
`--behavior-weak`	`#241337`		tint background

Indigo sits between teal and purple on the wheel — the brand color is literally the blend of the two things the product does. Composition endpoints (RAG + adapter) may use a teal→purple gradient to signal "both."

3.3 Neutrals (dark-first)¶

Token	Hex	Role
`--bg`	`#0E1117`	App background — "ground zero."
`--panel`	`#161B22`	Cards, sidebar, topbar.
`--panel-2`	`#1E2430`	Inset surfaces, inputs-on-panel, code blocks.
`--border`	`#2A313D`	Hairlines.
`--text`	`#E6E8EC`	Primary text.
`--muted`	`#9AA3B2`	Secondary text, labels.

3.4 Semantic¶

Token	Hex	Role
`--success` (`--green`)	`#22C55E`	Passed eval, healthy, ready. (Distinct from knowledge-teal.)
`--warning` (`--amber`)	`#F59E0B`	Pending, degraded, "needs a GPU."
`--danger` (`--red`)	`#EF4444`	Failed, destructive, revoke/delete.

Contrast: every text/background pair must clear WCAG AA (4.5:1); large text and non-text UI clear 3:1. --brand on --bg and all semantic colors on their -weak tints are AA-verified. Don't introduce a color outside this table without checking contrast.

4. Typography¶

An all-open-source, self-hostable stack (no tracking, works offline — matches the on-prem ethos).

Role	Family	Notes
Display / headings / wordmark	Space Grotesk	Geometric grotesk; engineered, characterful. Weights 500/700.
Body / UI	Inter	Neutral, dense, superb at 13–14px. Weights 400/500/600.
Mono / code / keys / IDs	JetBrains Mono	Already referenced in the console. Weight 400/500.

adp_* API keys, endpoint slugs, model IDs, correlation IDs, and code always render in mono.

Type scale (console, 14px base): 22 / 16 / 13 (uppercase, tracked, --muted) for h1/h2/h3 — this matches the existing console rhythm; the change is the families, not the sizes.

Fallbacks: keep Inter, system-ui, … and "JetBrains Mono", ui-monospace, … stacks so the UI is correct before webfonts load. Self-host the woff2 files; do not hot-link Google Fonts (an on-prem box may have no outbound internet).

5. Logo & mark¶

5.1 The mark — "the adapter ring"¶

A ring enclosing three connected nodes — your infrastructure boundary around a model composed in place. The three nodes are colored brand indigo · knowledge teal · behavior purple: the mark encodes the entire product (a base model adapted by knowledge + behavior, composed on one endpoint).

It is one flat SVG, legible at 16px (favicon) and on a billboard. This is the canonical source — generate every raster asset from it:

<svg width="64" height="64" viewBox="0 0 64 64" fill="none" xmlns="http://www.w3.org/2000/svg">
  <rect width="64" height="64" rx="16" fill="#0E1117"/>
  <circle cx="32" cy="32" r="20" stroke="#6366F1" stroke-width="3"/>
  <circle cx="32" cy="20" r="3.6" fill="#2DD4BF"/>   <!-- knowledge -->
  <circle cx="22" cy="40" r="3.6" fill="#6366F1"/>   <!-- brand -->
  <circle cx="42" cy="40" r="3.6" fill="#A855F7"/>   <!-- behavior -->
  <path d="M32 20 L22 40 M32 20 L42 40 M22 40 L42 40"
        stroke="#6366F1" stroke-width="2" stroke-opacity="0.5" stroke-linecap="round"/>
</svg>

For a transparent or light-surface variant, drop the <rect> and keep the ring + nodes.

5.2 The wordmark¶

Adapta set in Space Grotesk 700 (--text) as a single wordmark — matching Layout.svelte (<span class="wordmark">Adapta</span>), now with the brand face.

◐  Adapta

Lockup: mark + wordmark, horizontally, mark height = cap height × ~1.4, gap = ½ mark width.
Clear space: at least the ring's diameter on all sides.
Minimums: mark alone ≥ 16px; full lockup ≥ 120px wide.
Don'ts: don't recolor the nodes arbitrarily, don't stretch, don't add a drop shadow, don't set the wordmark in another typeface, don't put the dark-badge mark on a busy photo.

5.3 Favicon / touch icons¶

From the mark: favicon.svg (modern), favicon.ico 32×32 (legacy), apple-touch-icon.png 180×180, and a 512×512 maskable PWA icon. The dark rounded-rect badge variant is the default favicon so it reads on a white browser tab.

6. Application surfaces (where the brand lives)¶

This is the inventory the rollout targets. Details and acceptance live in TODO.md §B.

#	Surface	Files	What changes
1	Console tokens	`adapta/console/src/tokens.css` (new), `app.css`	Palette + type tokens become the single source; `app.css` consumes them.
2	Console shell	`components/Layout.svelte`, `views/Login.svelte`, `index.html`	Logo lockup, title/meta/OG, favicon links, theme-color, mode-colored cards.
3	Mode color-coding	project cards, model picker, choice grid	Knowledge=teal, Behavior=purple, composition=both.
4	Docs site	`mkdocs.yml`, `docs/stylesheets/brand.css` (new), `docs/assets/`	Indigo palette, brand fonts, logo, favicon.
5	API contract	`specs/openapi.yaml` (`info` block)	Branded title/description; Swagger UI logo + colors.
6	README	`README.md`	Branded header, tagline, logo, badges.
7	Package metadata	`pyproject.toml`	`description` aligned to the positioning line.
8	Voice pass	console microcopy, empty states, errors	Apply §2 — verbs not adverbs, two-mode framing, honest errors.
9	Email (future)	invite/transactional templates (none exist yet)	When built, inherit this system. Tracked as deferred.

Brand string SSOT. The product name/tagline currently lives in ~14 files. The rollout centralizes the console strings in one brand.ts constant and the server strings in adapta/config.py, so future renames are one edit, not a grep-and-replace.

7. Quick reference (copy/paste tokens)¶

:root {
  /* brand */
  --brand: #6366F1; --brand-weak: #1B1F3A; --brand-ink: #FFFFFF;
  --accent: var(--brand); --accent-weak: var(--brand-weak); /* back-compat aliases */
  /* modes */
  --knowledge: #2DD4BF; --knowledge-weak: #0E2E2A;
  --behavior:  #A855F7; --behavior-weak:  #241337;
  /* neutrals */
  --bg: #0E1117; --panel: #161B22; --panel-2: #1E2430;
  --border: #2A313D; --text: #E6E8EC; --muted: #9AA3B2;
  /* semantic */
  --green: #22C55E; --amber: #F59E0B; --red: #EF4444;
  /* type */
  --display: "Space Grotesk", "Inter", system-ui, sans-serif;
  --sans: "Inter", system-ui, -apple-system, "Segoe UI", Roboto, sans-serif;
  --mono: "JetBrains Mono", ui-monospace, "SF Mono", Menlo, Consolas, monospace;
  --radius: 10px;
}