The Bare Metal

On-device. Vertical-specific LLM weights. No external API calls — by architectural design. Runs on Apple Silicon (M4 Max / Ultra) or x86 / Proxmox infrastructure.

Two isolated databases: your sovereign documents (never leave your hardware) and a pre-loaded foundation knowledge corpus for your vertical (GPG-signed, atomically updated).

Immutable inference logs for every token generated. Cryptographically sequenced. Locally retained.

Traceable reasoning paths on every output. Every AI-assisted decision has a cryptographically sequenced record — satisfying ASIC, HREC, and SOC 2 audit requirements by construction.

MCP connectors across all four verticals: trading (Alpaca, Interactive Brokers, Databricks), research (arXiv, CrossRef, SSRN), clinical (PubMed, ClinicalTrials.gov, ICD-10/SNOMED CT), and software engineering (GitHub, Jira, Confluence). Plus universal connectors: Slack, Teams, Google Workspace, Microsoft 365, SQL databases. Systems of record are architecturally read-only — the agent produces the query, statement, or report; you execute it. Action tools are human-gated: the agent drafts, you decide.

MNPI-safe by architecture — zero tokens leave your perimeter. HREC/IRB-compatible data governance for research. SOC 2 and ISO 27001 audit evidence generated automatically. Compliance is a structural property, not a policy document.

Mac Studio M4 Max — 128 GB unified memory (546 GB/s): Recommended entry configuration. 128 GB is both the maximum M4 Max specification and the platform minimum for running a 70B vertical model with per-vertical LoRA adapters across all four verticals from a single node. Supports a professional team of 10–20 active users under sustained workload.

Mac Studio M3 Ultra — 192 GB to 512 GB unified memory (819 GB/s): Practice-group to organisation scale. Higher memory bandwidth (819 GB/s vs 546 GB/s) improves per-token throughput as well as concurrency. At 192 GB: supports 30–60 active users. At 512 GB — the current Mac Studio ceiling — the deployment is genuinely organisation-scale, with headroom for model ensembles and extended context workloads across concurrent teams.

Proxmox / NVIDIA — 80 GB+ VRAM: Software-only deployment for existing rack infrastructure. A100 80GB, H100 80GB, or RTX 6000 Ada (dual-card). CUDA acceleration delivers higher token throughput than unified memory at equivalent VRAM — suited to high-concurrency or batch-heavy workloads. Multi-GPU configurations (160 GB+) support larger models and concurrent agent pipelines.

Previous-generation Apple Silicon: Apple Silicon Macs with ≥128 GB unified memory — including M2 Ultra, M3 Max (128 GB), and M3 Ultra configurations — are supported as minimum-spec deployments. 128 GB is the absolute floor for running a 70B model with LoRA adapters at production quality.

Each role ships with pre-built Skills — structured prompt chains, tool sequences, and output templates for high-value tasks. Write custom Skills in plain YAML/Markdown and assign them to any role template via the admin UI. Zero skill execution leaves the perimeter.

All platform updates — Skills, MCP connectors, knowledge corpus refreshes, and new model weights — pass through cryptographic verification and automated regression testing before promotion to production. GPG-signed bundles only.

Licensed per node — not per seat, not per query, not per token. The licence is fixed per node regardless of how many users it serves. Capacity is determined by hardware; the licence never becomes the constraint. Add a node to expand capacity — not to unlock a higher pricing tier.