Interactive, real-time visualization of LLM inference internals. Submit a prompt, pick a model, and watch the generation pipeline animate token by token.
Status: Pre-implementation. Specification and execution plan complete. Code scaffolding has not begun. Domain: llm.trackr.live (not yet live)
LLM-Viz lets users submit a prompt to any of the major LLM vendors (OpenAI, Anthropic, Google, xAI, Meta-via-Groq/Together, Mistral, Groq, Together.ai, HuggingFace) and watch an animated visualization of the inference pipeline — tokenizer → embeddings → transformer layers → MoE routing (for MoE models) → logits → decoding — synchronized to the real token stream from the vendor's API.
It is designed for researchers, educators, developers, and enthusiasts who want intuition for how LLMs generate text without needing local GPU infrastructure.
- Not a model-hosting platform — inference happens on vendor APIs (BYOK).
- Not a tensor inspector — visualizations are illustrative, paced to real token-generation timing. (A future phase adds real tensor capture for open-weights models via a separate Python sidecar — see
docs/phase2.md.) - Not a fine-tuning or training tool.
| Layer | Tech |
|---|---|
| Backend | Laravel 13, PHP 8.4+, SQLite |
| Frontend | Inertia.js + React 19 + TypeScript, Vite, Tailwind CSS 3 |
| Visualization | Three.js (3D layer stack, token flow), D3.js (heatmaps, logits, MoE routing) |
| Realtime | Soketi (self-hosted, Pusher-protocol-compatible) |
| Auth | Laravel Sanctum + Socialite (Google / Microsoft / Facebook social login only) |
| Vendor HTTP | Guzzle 7 |
| Hosting | DreamHost VPS |
| Testing | Pest (PHP), React Testing Library, Cypress |
Full details in SPEC.md.
| Document | What's in it |
|---|---|
SPEC.md |
Authoritative specification: requirements, architecture, schema, acceptance criteria. |
docs/PLAN.md |
Execution roadmap across all phases. |
docs/phase1.md |
Detailed milestone breakdown for the launch product (14 milestones). |
docs/phase2.md |
Future: Python sidecar for tensor-accurate visualization of open-weights models. |
docs/phase3.md |
Future: multi-user collaboration, organizations, shared sessions. |
docs/phase4.md |
Future: education mode and side-by-side model comparison. |
The project is not scaffolded yet. The commands below describe the intended workflow once Milestone M1 of Phase 1 completes.
- PHP 8.4+
- Composer 2.x
- Node 20+
- SQLite 3
- (Optional, for realtime) Soketi via Docker
git clone git@github.com:CyberSecDef/llm.trackr.live.git
cd llm.trackr.live
composer install
npm install
cp .env.example .env
php artisan key:generate
touch database/database.sqlite
php artisan migrate --seed# Terminal 1 — Laravel
php artisan serve
# Terminal 2 — Vite
npm run dev
# Terminal 3 — Queue worker
php artisan queue:work
# Terminal 4 — Soketi (or use Docker)
soketi startVisit http://localhost:8000, sign in with a social provider configured in .env, add an API key for any supported vendor, and create a thread.
php artisan user:promote you@example.com.
├── README.md ← you are here
├── SPEC.md ← authoritative spec
├── docs/
│ ├── PLAN.md ← cross-phase execution plan
│ ├── phase1.md ← active phase, detailed milestones
│ ├── phase2.md ← future: Python sidecar
│ ├── phase3.md ← future: collaboration
│ └── phase4.md ← future: education + comparison
└── ... (Laravel scaffold to be added in M1)
The project is in pre-implementation planning; external code contributions aren't being accepted yet. Issues, design feedback, and discussion are welcome via GitHub Issues. A CONTRIBUTING.md will appear before the M1 scaffold lands.
Licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE for the full text.
The AGPL is a strong copyleft license: if you run a modified version of LLM-Viz as a network service, you must make your modifications available to users of that service under the same license. This is intentional — it lets others self-host or fork while preventing closed-source SaaS clones.