# Velora by KMATS — Complete System Knowledge Compendium

> Complete technical documentation, hardware architecture, benchmark figures, inference mechanics, and legal warranties for Velora by KMATS Labs.

---

## 1. System Overview & Core Proposition

Velora is a purpose-built, portable physical drive containing a turnkey offline AI assistant. Unlike cloud AI subscriptions (OpenAI ChatGPT, Anthropic Claude, Microsoft Copilot, Google Gemini) which require ongoing monthly payments, user authentication, and transmitting user prompts to remote data centers, Velora runs 100% locally on the user's computer with zero network connectivity.

### Key Value Pillars
1. **Zero Recurring Fees**: One-time hardware purchase from ₹1,500. No monthly subscriptions, no API credit consumption, no rate limits.
2. **True Air-Gapped Privacy**: Operates fully with the computer's Wi-Fi and Ethernet disconnected. No telemetry, no logs sent to remote servers, no user accounts.
3. **Turnkey Portability**: Zero installation on host machine. Plug into any Windows 10/11 PC and launch the pre-compiled executable directly from the drive.
4. **Industrial-Grade Flash**: 128 GB NVMe storage capable of 1,050 MB/s sustained sequential reads with graphene passive cooling to prevent thermal throttling during continuous memory-mapped tensor reads.

---

## 2. Hardware Architecture & Thermal Design

- **Storage Media**: High-endurance 128 GB 3D TLC NVMe NAND Flash
- **Controller Interface**: PCIe Gen3 x4 bridged via Realtek / ASMedia high-speed USB 3.2 Gen 2 controller
- **Physical Enclosure**: CNC-machined 6063 Anodized Aluminum casing
- **Thermal Dissipation**: Direct-contact graphene heat pad transferring thermal energy from the NAND controller to the outer aluminum shell (max operating temperature < 48°C under continuous load)
- **Power Consumption**: 2.1W idle, 3.8W peak during model weight mmap initial load; powered directly via standard 5V USB bus without external power brick
- **Connectors**: Dual integrated USB-C and USB-A 3.2 connectors

---

## 3. Software Stack & Inference Pipeline

1. **Inference Engine**: Custom static build of `llama.cpp` compiled with AVX, AVX2, FMA, and OpenMP vector extensions for x86_64 and NEON for ARM64.
2. **Acceleration Backends**:
   - Direct Vulkan compute shaders for universal acceleration on NVIDIA, AMD, and Intel GPUs.
   - Native CUDA backend for NVIDIA GeForce, Quadro, and RTX GPUs.
   - Multi-threaded CPU fallback with automatic core affinity binding.
3. **Weight Format**: GGUF (GPT-Generated Unified Format) with Q4_K_M (4-bit medium k-quant) and Q8_0 quantizations for optimal perplexity-to-memory footprint.
4. **Multi-Modal Capabilities**:
   - Local Whisper (`whisper.cpp`) for zero-cloud audio transcription.
   - Offline Vector Search (`ChromaDB` / `hnswlib` embedded) for local document retrieval and RAG (Retrieval-Augmented Generation) without internet.
5. **Local Web Interface**: Light-footprint, sandboxed HTTP server serving a modern responsive web client locally at `http://127.0.0.1:8080`.

---

## 4. Hardware Editions, Pricing & Benchmarks

| Edition | Price (INR) | Recommended Host Specs | Model Weights | Expected Generation Speed |
|:---|:---|:---|:---|:---|
| **Core** | ₹1,500 | Intel i5 / Ryzen 5, 8 GB RAM | 3B Instruct (Q4_K_M) | 28–45 tokens/sec (CPU), 65–95 tokens/sec (GPU) |
| **Pro** | ₹2,999 | Intel i7 / Ryzen 7, 16 GB RAM | 7B/8B Instruct (Q4_K_M) | 16–26 tokens/sec (CPU), 40–70 tokens/sec (GPU) |
| **Prime** | ₹5,499 | Intel i7/i9, 16–32 GB RAM | 14B High-Density (Q4_K_M) | 8–15 tokens/sec (CPU), 25–45 tokens/sec (GPU) |
| **Ultra** | ₹9,999 | Intel i9 / Ryzen 9, 32 GB+ RAM | 32B Precision (Q4_K_M/Q8) | 4–9 tokens/sec (CPU), 18–30 tokens/sec (GPU) |

---

## 5. Privacy, Air-Gapping & Security Model

- **No Network Permissions**: The Velora runtime executable does not bind to external interfaces (0.0.0.0). It strictly binds to `127.0.0.1` (localhost loopback).
- **No Third-Party Cookies or Analytics**: The web interface does not load external CDN scripts, fonts, analytics, or trackers. All assets, fonts, and scripts reside locally on the drive.
- **Hardware Isolation**: When the drive is ejected, zero application files or registry keys remain on the host computer.

---

## 6. DIY Blueprint vs. Velora Turnkey Comparison

For developers and hobbyists, KMATS provides the full DIY guide at `https://velora.kmats.in/nerd/`.

1. **DIY Route**: Requires purchasing a fast flash drive, testing flash integrity with ValiDrive/H2testw, compiling `llama.cpp` for Windows with CMake/MSVC, pulling GGUF models from Hugging Face, setting up chat templates, configuring llama-server, and bundling portable runtimes.
2. **Velora Turnkey Route**: Delivers an industrial 1050 MB/s NVMe drive with graphene thermal management, pre-compiled zero-dependency binaries, validated models with optimized quantizations, a 1-year hardware replacement warranty, and 7-day money-back guarantee.

---

## 7. Company & Manufacturer Information

- **Company**: KMATS Labs
- **Location**: Bengaluru, Karnataka, India
- **Support & Inquiries**: contact@kmats.in / security@kmats.in
- **Website**: https://velora.kmats.in