Tactical AI Workbench & Hardware Key Enclave.
Explore the dual operational vectors of Project Howitzer: the high-performance native desktop creation suite on Apple Silicon, and the mobile hardware enclave companion that prevents API key leaks.
Universal Model Freedom
Never be constrained by closed platforms. Howitzer treats all frontier models and open-weight checkpoints as first-class citizens.
Frontier Cloud APIs
Direct streaming connections to OpenAI (GPT-4o, o1), Anthropic (Claude 3.5 Sonnet, Haiku), Google (Gemini 3.5 Flash), and Mistral (Large 2, Codestral). Sub-millisecond dispatch with automatic failover.
1-Click Curated Local GGUFs
Download and run optimized GGUF weights directly on your local GPU with zero command-line configuration. Pre-quantized Qwen 2.5 Coder, DeepSeek R1 Distill, Llama 3.3, and Mistral NeMo.
Custom Weights & External Paths
Load any custom `.gguf` or MLX model file from your filesystem. Configure context lengths, GPU offloading layers, and temperature seeds with instant parameter reloading.
The Foundry Creation Suite
Take complete ownership of your model lifecycle. Generate instruction datasets from raw seed documents, fine-tune custom adapters on Apple Silicon, and package for distribution.
Agentic Dataset Synthesis
Autonomous multi-agent synthesis loops. Feed in codebase folders or technical PDFs; agents generate diverse reasoning pairs with automated adversarial critique and schema validation.
- Rejection:Automated AST & Hallucination Pass
- Formats:Alpaca, ChatML, ShareGPT, JSONL
Desktop LoRA / QLoRA Tuning
Hardware-accelerated training using Apple Silicon Metal shaders. Fine-tune 3B to 8B models without renting remote cloud infrastructure or configuring CUDA drivers.
- Telemetry:Real-Time Loss & Perplexity Charts
- Hot-Swap:Instant Adapter Switching in Memory
1-Click GGUF Quantization
Convert trained checkpoints to edge-ready GGUF (Q4_K_M, Q5_K_M, Q8_0) or MLX formats with automated degradation assertions to protect downstream accuracy.
- Compression:4-bit & 8-bit Integer Quantization
- Export:Standalone Portable Model Files
Live Canary Range
Spin up an instant local OpenAI-compatible endpoint. Conduct side-by-side verification: Base Model vs. Fine-Tuned LoRA vs. Frontier Baseline, backed by automated regression tests.
- Benchmarking:Zero-Forgetting Regression Suites
- Serving:Instant Local HTTP /v1/chat/completions
The Five Execution Batteries
Pure C++ Qt 6.5.3 desktop workstation. Instantaneous boot, sub-millisecond dispatch, and deterministic memory management.
[DF] DirectFire
Precision API payload inspector. Real-time latency charts, HTTP/2 multiplexing, custom SSL pinning, and payload body formatters.
[AR] Arsenal
Dynamic prompt template studio. Bracket mustache variable expansion, strict JSON AST schema validation, and hyperparameter dialing.
[BT] Battery
Multi-stage pipeline sequencer. Chain generation stages with regex and JSONPath extraction rules, running in sequence or parallel.
[FR] Firing Range
Tri-model shootout arena. Compare 3 frontier or local models side-by-side with unified token diffing and automated arbiter consensus.
[MZ] Magazine
Hardware enclave monitor and environment secret vault. Masked credential tables, multi-environment toggles, and shell integration.
Native Howitzer CLI
All desktop capabilities are accessible directly from your terminal. Pipe standard Unix streams, assert model outputs in CI/CD, and inject secrets securely into your shell.
cat server.log | howitzer run qwen-2.5-coder "Extract anomalous 500 errors into JSON"
# 2. Automated multi-model shootout assertion in CI/CD
howitzer shootout --models gpt-4o,claude-3.5-sonnet,local-qwen --diff --assert-schema ./schema.json
# 3. Headless dataset synthesis from directory of documentation
howitzer foundry synth --input ./docs/ --output ./eval_dataset.jsonl --samples 500
# 4. Inject encrypted secrets directly into shell environment
eval $(howitzer env export --env prod)