AI Launchpad
Open a local model or launch an online service. Local buttons require LM Studio and model identifiers that match the installed Aegis Box manifest; cloud services require internet access.
Local Models // Aegis Box ABL-SOVR Stack
Cloud Chat AIs // Online
API & Agent Frameworks
Rig: Hardware Inventory
Primary // Aegis Box ABL-SOVR Edition
AI Processor NEW
AMD Ryzen AI Max+ 395 with integrated Radeon graphics and NPU. The large unified-memory pool reduces discrete-VRAM constraints for local inference.
System Memory NEW
128GB LPDDR5X unified memory — suitable for quantized 70B-class models and selected larger MoE workloads, depending on context and offload settings.
Form Factor
Rugged Tactical Exoskeleton — desk to field-ready in seconds.
Display Out
Portable 8K Field Monitor — immersive data visualization anywhere.
Input Device
GravaStar Mechanical Keyboard — ergonomic tactical input.
Connectivity
Offline-capable by default. Optional authenticated networking can be enabled selectively for approved services and remote access.
Extended Field Kit
Tele-Op Agent
Experimental remote-mobility integration using Android, approved network access, live video, and operator stop controls. Range depends on hardware, network conditions, and local rules.
Vision System
Ricoh Theta Z1 — 360° environmental capture for digital twin and VR work.
Ghost HUD
Meta Ray-Ban glasses for hands-free audio and capture; XREAL or Viture displays for private visual overlays and field interfaces.
Power System
DJI Power 500/1000 + Zignes Solar Panel — off-grid autonomy for field operations.
Connectivity
Portable network switch + optional Starlink for selectively connected deployments.
Projector Array
4× Salange HY300 Pro — multi-wall immersive projection. See PROJECTOR module.
Toolkit: Software
AI Inference Interfaces
LM Studio
Primary desktop interface for downloading, configuring, loading, and serving local models. Current builds also support local APIs and MCP integrations.
Ollama
Lightweight always-on background engine for serving models via API.
Jan
Polished user-friendly interface focused on the chat experience.
Oobabooga WebUI
Advanced experimental web interface for model loading, extensions, evaluation, and custom inference workflows.
Creative & Specialist Tools
ComfyUI
Node-based local image and media workflow engine. AMD compatibility depends on the selected runtime and model stack.
SillyTavern
Character-based AI roleplay and creative writing frontend.
Buzz / Whisper
Offline audio-to-text transcription service using the Whisper model.
AnythingLLM
Command center for RAG — chat with your own documents, entirely locally.
Tasker (Android)
Android automation bridge for hands-free capture, notifications, and approved local-AI workflows.
Immersive & Projection Tools
TouchDesigner
Real-time visual programming for projection mapping and generative scenes.
Resolume Arena
VJ software for multi-screen mapping, edge blending, and live performance.
MadMapper
Dedicated projection mapping with keystone correction and sync.
Minds: Model Library
Aegis Box // Current Local Stack
Nous Hermes-4 70B CURRENT
Current large local generalist for analysis, drafting, document work, and instruction following. Quantized GGUF deployment on unified memory.
Qwen3-Coder-Next CURRENT
Primary local coding model for architecture, debugging, repository analysis, and long-context software tasks.
GLM-4.5-Air CURRENT
Large generalist and coding model used for complex local workflows when additional reasoning depth is useful.
Local Image Stack MEDIA
ComfyUI-based image generation and transformation. Local operation improves control; model licenses and output review still apply.
Heavyweight Generalists // Extended Arsenal
WizardLM-2 8×22B (MoE)
Experimental heavyweight Mixture-of-Experts model. Memory use and speed depend on quantization, active experts, and context length.
DeepSeek R1 Distill 70B
Reasoning-focused distilled model for deliberate analysis, mathematics, and technical problem solving.
Qwen3 4B Sky-High Hermes
Compact always-on assistant for fast local interactions, lightweight automation, and constrained hardware profiles.
Vision Specialists
Qwen3-VL 8B Instruct
Primary local vision-language model for screenshots, documents, interfaces, and image-grounded questions.
MiniCPM-V
Efficient secondary vision model for fast comparison, extraction, and edge-oriented visual tasks.
Experimental & Creative Specialists
Creative Writing Specialist
Isolated narrative and roleplay model. Outputs remain drafts and receive the same privacy, copyright, and safety review as other models.
NeuralDaredevil 8B — Experimental
Modified-alignment research model. Keep isolated from tools and sensitive data; reduced refusal behavior does not improve accuracy or judgment.
Dolphin Llama 3 — Experimental
Creative multilingual model with modified alignment behavior. Use in permissioned local workflows with logging and human review.
Finance Hub
Trading Platforms & APIs
Robinhood
Retail brokerage interface for manual monitoring and research. No live trading integration is enabled by default.
Alpaca Markets
API-oriented brokerage platform suitable for paper-trading experiments and controlled automation.
Core Concepts
Momentum Strategy
A simple research baseline based on recent price movement. It is not a prediction guarantee or investment recommendation.
The Two Dials
Experimental controls for trade frequency and position value. Hard limits and independent risk controls are required.
Paper Trading
Required first stage for validating data, orders, limits, failures, and reporting before any live-capital use.
Robotics Hub
Tele-Op // Aegis Box Integration
Segway S2 Agent
Experimental remote-mobility integration using Android, encrypted networking, live video, edge vision, and operator stop controls. Operate only where authorized and within platform limits.
Tailscale VPN Mesh
Encrypted WireGuard mesh with peer-to-peer paths when possible and relay fallback when required. Access still depends on the configured coordination and network services.
FPV Drone
Aerial capture and visualization experiments. Flights require operator control, applicable authorization, and compliance with local aviation rules.
Cloud & Local Frameworks
Gemini Robotics Research
Reference track for current vision-language-action and embodied-AI research; not part of the offline core stack.
ROS 2
Widely used open robotics middleware for messages, sensors, control, visualization, and multi-node integration.
NVIDIA Isaac Sim
GPU-oriented robotics simulation and synthetic-data environment; optional external tool rather than a requirement for ABL_SOVR.
Project Log: Archive
Aegis Box — ABL-SOVR Hardware Build
Ryzen AI Max+ 395, 128GB LPDDR5X. Tactical exoskeleton. Fully sovereign local inference stack.
BG3 Stereo VR Mod — Active
Side-by-side stereo injected into Baldur's Gate 3 without official VR support. Three tests logged.
Projector // Glitch Room — In Dev
4× HY300 Pro. U-shaped 3-wall panoramic display. Glitch transitions, portal effects, 360° video.
Tele-Op Mobility Agent — Experimental
Experimental Segway-based mobility interface with Pixel vision, encrypted remote access, and operator stop controls. Range and use depend on platform limits, network conditions, and authorization.
Software Stack Deep Dive
LM Studio, Ollama, SillyTavern, ComfyUI configured on Aegis Box. lmstudio:// deep links active.
Model Curation — ABL-SOVR Manifest
Current manifest includes Nous Hermes-4 70B, Qwen3-Coder-Next, GLM-4.5-Air, Qwen3-VL 8B, DeepSeek R1 Distill 70B, and experimental specialist models.
Advanced AI Concepts Reference
RAG, MoE, Quantization, Knowledge Distillation, Abliteration defined. See Codex.
Project Ideation Archive
Workshop Guardian, AI Game Companion, Glitch Room, stereo injection pipeline, tele-op robotics.
Codex: Project Glossary
- ABL_SOVR
- The name of the local-first rig and software stack. ABL_SOVR emphasizes operator ownership, offline capability, model transparency, and configurable boundaries rather than dependence on a single cloud provider.
- Abliteration
- A research term for interventions that reduce or remove model refusal behavior. These interventions can also remove useful safeguards and do not improve factual accuracy; experimental models should remain isolated, permissioned, logged, and human-reviewed.
- GGUF
- A model file format used by llama.cpp-compatible runtimes. GGUF supports metadata and quantized weights for efficient local inference across CPU, GPU, and unified-memory systems.
- Quantization (Q4_K_M)
- Quantization reduces model memory and compute requirements by representing weights at lower precision. Q4_K_M is a common quality/size tradeoff, not a universal best choice; evaluate by model, task, context, and hardware.
- RAG (Retrieval-Augmented Generation)
- Retrieval-Augmented Generation supplies selected external documents or records as context at query time. It can improve grounding, but retrieved sources still require provenance, permissions, freshness checks, and citation review.
- MoE (Mixture-of-Experts)
- Mixture-of-Experts models route tokens through selected expert components rather than activating every parameter for every token. Memory, speed, and quality depend on architecture, quantization, routing, and runtime support.
- Knowledge Distillation
- Knowledge distillation trains a smaller student model using outputs or signals from a larger teacher. Distilled models can be efficient, but they do not automatically inherit every capability or reliability property of the teacher.
- DPO-Healed
- Direct Preference Optimization tunes model behavior using preferred and rejected response pairs. Community labels such as “healed” are informal and should not be treated as standardized safety or quality certifications.
- Tailscale VPN Mesh
- An encrypted WireGuard-based overlay network for connecting approved devices. Connections are peer-to-peer when possible and may use relay infrastructure depending on network conditions and configuration.
- Glitch Room
- Multi-projector immersive environment treating a room as a corrupted simulation. Scenes load through glitch transitions, portal tears, and scanline wipe effects.
- lmstudio://
- A local deep-link convention used by this dashboard to request model loading in LM Studio. It works only when LM Studio supports the handler and the link’s model identifier matches an installed model.