Writes in your voice
Prompted voice matching gets you "sounds like me." A model fine-tuned on your actual writing gets you "is me." Drafts ship unchanged.
Your chief of staff
Every founder, builder, and exec yearns for a great chief of staff — intelligent, tireless, and tuned to their style. Better than an entry-level grad, 100× cheaper, and available 24/7 via fine tuning.
Why Samurai
Takes action
Agentic write privileges. Send the email. Book the meeting. Update the doc. You approve the action, Samurai executes it.
Verifiable memory
Programmatically enforced memory. The model doesn't just remember — it knows. Memory is 0 or 1, not probabilistic. Zero hallucinations on facts about your own life.
Retrained weekly
Your life evolves. The model evolves with it. New patterns are folded into the weights every week.
How it works
Three stages, fully automated.
Stage 1
Cluster
Robot Dojo scans your imported data — emails, messages, calendar — and clusters your writing into natural voice types. Professional outbound. Casual texts. Internal notes. Each gets a training set.
Stage 2
Train in the cloud
Your data is encrypted and sent to an isolated AWS GPU pod. A LoRA adapter is trained on your specific patterns. No multi-tenancy, no shared memory. Training data is purged immediately afterward.
Stage 3
Inference via private API
Your fine-tuned model is hosted on our infrastructure and exposed to your machine over a private API. No 128GB of RAM required. Raw data never leaves your machine — only prompts and retrieved context.
Infrastructure you can trust
Your data is never at rest on our servers.
Isolated AWS training pods
Training runs on isolated AWS GPU instances. Each customer gets a dedicated pod — no multi-tenancy, no shared memory, no data leakage between users.
Encrypted end-to-end
Your data is encrypted before it leaves your machine. It stays encrypted on AWS until the training pod decrypts it in memory. After training completes, all source data is purged.
You own the weights
Your LoRA adapter weights are yours. Export them in standard formats. Cancel anytime — we hand you the weights on the way out.
Data lifecycle
Training data exists in AWS only during active training (typically a few hours per week). Afterward it's deleted. Inference runs against your hosted model — the prompts and context you send are not stored or used for future training.
Black Belt vs Samurai
Black Belt is the intelligence layer. Samurai is the agent layer — the 20% that writes in your voice and acts on your behalf.
| Black Belt | Samurai | |
|---|---|---|
| Drafting | Prompted voice matching (sounds like you) | Fine-tuned weights (is you) |
| Actions | Read-only intel | Write privileges (with your approval) |
| Timeline | Auto-compiled timeline | Cryptographically signed + verifiable |
| Learning | Context window per session | Model retrains weekly |
| Voice types | Manual selection | Auto-detected from patterns |
| Ownership | Foundation model via API | Your weights, exportable |
Samurai is coming
Samurai is available first to Black Belt subscribers.
Add your Black Belt account ID below.
Miyagi will ping you in-app when your slot opens.