Infrastructure

AI Lab

Private infrastructure for finance and investment banking work: several models with distinct roles, strict privacy classification, evidence before narrative, and human approval before anything is published.

Architecture

How the stack is arranged

Capture
Voice input
Text and documents
Short-form capture channel
Collaboration
Planning and adjudication
Implementation lane
Adversarial review lane
Local research interface
Memory
Multi-domain knowledge graphs
Checksummed evidence archive
Source-integrity acquisition
Private knowledge bases
Compute
On-device hardware
Private model runtime
Local embeddings
Control
Privacy classification
Security monitoring
Human approval gates
Workflow automation
Work ledger and health checks
Philosophy

Design principles

Classify before you transmit

Material is sorted by sensitivity before any model sees it. Open research may use cloud capacity deliberately. Sensitive excerpts are minimised and logged. Restricted records stay on the machine. Employer and client work enters only as a human-written brief, never as raw files.

Evidence before narrative

Public sources are fetched under a fixed institutional identity, with robots checks, hard budgets, and fail-closed stops. Bodies are checksummed into an evidence archive before they inform analysis. Knowledge graphs keep durable context; they do not invent facts that the archive cannot support.

Several models, one human gate

Planning, implementation, and review are separate lanes, often different models, with cross-checks between them. Cost is deliberate. Publication, valuation conclusions, and client-facing judgement remain human. No tool is allowed to claim work it did not do.

Capabilities

What the stack is for

Research and synthesis

First-pass reading of public regulatory, market, and policy material; structured notes; comparison across sources. Primary evidence is retained with provenance. Drafts are challenged before they are treated as ready.

Build and verification

Code, collectors, and internal tooling are built in an implementation lane and checked in a separate review lane. Tests and health checks are part of the loop, not an afterthought.

Controlled automation

Repetitive handoffs and scheduled checks run under the same privacy and security rules as interactive work. Remote access is closed by default. Escalation to cloud capacity is explicit, not ambient.

How work actually runs

Capture arrives as voice, text, or a short message. Anything larger than a quick answer becomes a ticket with a privacy class, an objective, and acceptance criteria. Planning and adjudication decide the approach; an implementation lane does the mechanical work; a second model reviews before the gate. Durable knowledge sits in multi-domain graphs. Public-source bodies sit in a checksummed archive and only enter product or publication surfaces when rights and review allow. On-device models handle work that must not leave the machine. Cloud use is cost-aware and lane-bound. A work ledger and periodic health checks show what ran, what failed, and what still needs a person. Nothing publishes itself.