Tech Tweakers LogoTech Tweakers

Our Active Projects

From protocols to runtimes to trained models — these are the projects we build and use in production.

Protocol

XCT — Execution Control Transfer

An architectural protocol that keeps language models from granting themselves executive authority. XCT inverts the sovereignty model: models propose, but don't execute. Control remains with deterministic tools. Born from production experience with high-risk AI deployments.

Protocol Principles

  • No action without explicit tool invocation
  • One step per iteration — controlled loop
  • Errors are first-class control signals
  • System has veto power over any proposal
  • Ambiguity resolves to minimal action

Why XCT Exists

XCT is not a prompt style or agent framework — it's an architectural protocol. Traditional AI integrations let models decide what to do and when. XCT inverts this: the model proposes, the system validates, deterministic tools execute. Authority remains external by design.

AI SafetyProduction ProtocolSovereignty InversionApache 2.0
Runtime

Polaris Core

Ultra-optimized C++ binding for llama.cpp with first-class Python integration. Built for maximum-performance local inference with streaming, XCT protocol support, and intelligent batch management.

Features

  • GGUF model support (Qwen, Mistral, LLaMA, Gemma)
  • GPU acceleration via CUDA (CPU/Metal also supported)
  • Token streaming through Python callbacks
  • JSON early-stop for XCT protocol (~30% time savings)
  • Intelligent batch backoff with automatic retry
  • Thread-safe architecture with GIL management

Performance (Qwen3-4B · RTX 4090)

~15ms
Prefill (256 tokens)
~2ms
Per token decode
~9GB
VRAM usage
30%
Savings with early-stop
C++Pythonpybind11llama.cppCUDAApache 2.0
AI Assistant

Polaris v2

Multi-modal AI assistant with conversational AI, voice synthesis, document processing, and multi-platform connectivity. Customizable chat deployable on local infrastructure or cloud — your models, your data, your rules.

Capabilities

Conversational AI
Dual-layer memory with semantic search
Voice I/O
TTS (ElevenLabs, Coqui) + STT (Whisper)
Document Processing
PDF vectorization + semantic search
Multi-Platform
REST API, Telegram, Web UI, Webhooks

Infrastructure Options

Local
llama.cpp + your GPU
Cloud
Groq API + managed infra

Architecture

  • FastAPI backend with JWT authentication
  • MongoDB persistence + ChromaDB vector store
  • LangChain context management
  • Dynamic model selection (local or cloud)
  • Prometheus metrics and observability
  • Docker Compose deployment
PythonFastAPIMongoDBChromaDBDockerllama.cppGroq
Automation Platform

Polaris v3

Automation platform powered by the XCT Protocol. Models propose actions, the system validates and controls execution through deterministic tools. Built for infrastructure orchestration, deployment pipelines, and compliance-sensitive environments where reliability matters more than speed.

XCT Execution Flow

1User defines task and context
2Model proposes exactly one action
3System validates the proposal
4Deterministic tool executes or rejects
5Result feeds back as signal for next iteration

Tool Ecosystem

Shell Execution
Commands with timeout and streaming
Docker & EKS
Container and cluster orchestration
Document Validation
Critical document verification flows
Blockchain Trading
Deterministic trade execution on-chain

Architecture

  • Clean Architecture with layered separation
  • FastAPI REST + WebSocket executor
  • RAG with ChromaDB for contextual retrieval
  • MongoDB dual-layer memory (short + long term)
  • Local inference (Polaris Core) or Groq Cloud
  • Hot-reload configuration without restarts
  • Prometheus metrics and observability
PythonFastAPIWebSocketMongoDBChromaDBXCT ProtocolMIT
Trained Model

XCT-Qwen3-4B

An execution-oriented language model fine-tuned with LoRA to operate under the XCT Protocol. Not a conversational assistant — a constrained execution engine for controlled, deterministic systems.

Model Specs

Base ModelQwen3-4B
Parameters~4B
TrainingLoRA r=8 · 5 epochs · 150 steps
Dataset238 curated examples
ArchitectureDecoder-only Transformer

Quantized Variants (GGUF)

Q2_K
1.67 GB
Q5_K
2.7 GB · Recommended

Designed For

  • Deterministic execution agents
  • Infrastructure orchestration (K8s, CI/CD)
  • Deployment automation pipelines
  • Tool-driven workflows
  • Compliance-sensitive environments

Execution Model

1System provides context and instruction
2Model proposes a response or action
3System validates and executes or rejects
LoRAQwen3GGUFXCT ProtocolApache 2.0