Senon AI
An AI ecosystem spanning language, vision, computer-use, voice, agent, and orchestration systems.
- Thread-Ripper 1.2 and 3.5
- Capybara Vision
- Platypus
- Capybara STT
- Capybara TTS
- VERONICA
- AI councils and hierarchical agents
Research
Models, agents, and the systems that make them useful.
My research work focuses on multimodal intelligence, computer use, long-horizon agent execution, and efficient model runtimes across CPU and GPU hardware.
CPU and GPU inference optimisation
Research into faster, more memory-efficient model execution across heterogeneous hardware.
- Matrix-operation optimisation
- Memory-bandwidth efficiency
- Cache-aware execution
- Low-level Rust components
- Scheduling and throughput
Senon AI model systems
A specialised family, not one general-purpose wrapper.
The Thread-Ripper models are sparse mixture-of-experts architectures fine-tuned from open-source base models. Each surrounding model holds a narrower role inside the wider ecosystem.
Thread-Ripper 1.2
The primary model in the Thread-Ripper family: a sparse mixture-of-experts architecture at 1.1 trillion total parameters with 32 billion active per token, across a one-million-token context window, fine-tuned from an open-source base for reasoning, tool use, agent execution, software development, and organisational workflows.
- Sparse MoE: 1.1T total / 32B active
- 1M-token context
- Tool execution
- Long-horizon agent tasks
Thread-Ripper 3.5
The advanced multimodal Thread-Ripper model, sparse mixture-of-experts at 1.6 trillion total parameters with 49 billion active per token across the same one-million-token context, extended to understand several information types including native video rather than depending entirely on extracted frames or transcripts.
- Sparse MoE: 1.6T total / 49B active
- 1M-token context
- Native video understanding
- Hierarchical agent spawning
Capybara Vision
A hybrid approach to computer use that does not reason about pixel positions at all, pairing a language model with a high-precision OCR pipeline to read an interface and work out what it means.
- Language model and OCR pipeline
- Interface reading
- No pixel-position dependency
- Multi-step computer tasks
Platypus
A conventional computer-use model that understands pixels directly, grounding its actions in the image itself for the work where spatial position is what matters.
- Pixel-level grounding
- Precise interface interaction
- High-accuracy input control
- Workflow automation
Capybara STT
A 1.75-billion-parameter speech-recognition model trained in-house, keeping voice input transcription inside Senon Solutions' infrastructure rather than routing audio to a hosted transcription API.
- 1.75B parameters
- Conversational transcription
- Voice command input
- VERONICA integration
Capybara TTS
A 3-billion-parameter speech-synthesis model providing VERONICA's spoken output for conversation, application feedback, and voice-controlled workflows.
- 3B parameters
- Conversational speech
- Application feedback
- Voice workflows
Research from model behaviour to execution efficiency.
The programme connects multimodal intelligence, computer use, agent coordination, runtime performance, memory efficiency, and scheduling into one applied research direction.