Data Sovereignty Architecture

Security & Architectural Specifications

FineTuneMyAI is engineered specifically for regulated industries, proprietary IP, and security-first engineering teams.

Operational Boundaries: Local Training vs. Hosted Inference

Mode 1: Free Local Fine-Tuning ($0)

Execution takes place entirely on your physical workstation (Apple Silicon, NVIDIA CUDA, or Windows). Model weights, training datasets, raw text, embeddings, and LoRA/QLoRA adapter checkpoints remain strictly on local storage. No training data is ever transmitted to the cloud.

✓ On-device compute & storage
✓ Zero weight/corpus uploads
✓ Offline training continuity
Mode 2: Paid Hosted Inference (Cloud)

When you explicitly choose to deploy a trained model for cloud serving, the exported adapter weights are securely transmitted via TLS 1.3 to FineTuneMyAI managed GPU clusters. Compute is billed per GPU second consumed with isolated container runtimes and OpenAI-compatible API authentication.

✓ Explicit user-initiated export only
✓ Isolated tenant container runtimes
✓ TLS 1.3 encrypted data in transit

What Remains On-Device in Local Mode

Base model weights & quantized NF4 files
Raw documents, CSVs, and JSONL corpora
Training prompt and completion pairs
Tokenizer cache and vocabulary files
LoRA / QLoRA adapter weights and checkpoints
Optimizer states and random number generator seeds
Local vector embeddings and RAG indexes
Absolute local filesystem directory paths

Cryptographic Device Pairing Architecture

Device registration establishes a cryptographically authenticated channel between your browser control plane and your local daemon:

1. Ephemeral One-Time Code

The web control plane generates a time-limited 6-digit pairing code linked to your authenticated user account. Codes expire automatically after 10 minutes and enforce strict rate limiting.

2. Local Ed25519 Key Generation

When paired, the local agent generates an asymmetric Ed25519 keypair entirely on-device. The private key never leaves the workstation and is stored in secure OS storage (macOS Keychain, Linux Secret Service, or Windows Credential Manager).

3. Mutual Verification & Scoped Bearer Tokens

The server validates the device public key and issues a high-entropy 256-bit bearer token, storing only its SHA-256 hash. Signed device requests use Ed25519 verification together with timestamps and replay-resistant nonces; the bearer token is scoped to the paired device credential.

Allowlisted Lifecycle Commands (No Arbitrary Shell Execution)

The FineTuneMyAI control plane does not expose arbitrary shell commands or remote execution interpreters. Communication is strictly restricted to typed, allowlisted ML lifecycle operations (TRAIN_START, TRAIN_STOP, EVALUATE_START, INDEX_START). Any command outside this validated manifest is rejected at the daemon boundary.