Security and compliance
Security is your bedrock.
Private capital firms can't take chances on security, accuracy, or compliance. Bloxii is built on a simple discipline: numbers are data, never guessed; every output cites the transaction behind it; and a person on your team approves before anything leaves the building. Below is the full architecture, isolation, audit, residency, and the controls auditors and regulators expect. A security pack is ready for your IT review.
Trust Centre: the security packPassword protected · talk to a founder for accessWho can reach your data
Your data, controlled and accounted for.
No training on your records and no quiet access. Your data is stored in the UK, runs behind one controlled gateway, and is isolated from every other firm. Who can reach it, and how that is logged, is written down in the Trust Centre. A person on your team signs off on every output.
Access, stated plainly
Bloxii’s platform administration has no designed read path to your content, and every action it takes lands on the append-only audit trail. Infrastructure access is held by a small number of named engineers, is audited, and is described plainly in the Trust Centre rather than implied away.
UK data residency on AWS London
Your application and stored data are hosted in the United Kingdom, on AWS London (eu-west-2), region placement enforced in CI. Model processing depends on the models enabled for your firm; where you require it, a firm can be given a UK-only model selection so no prompt or output leaves the UK.
One controlled AI gateway
Every model and retrieval call passes through the Bloxii Router: guardrails, prompt management, safety, rate limiting, caching, cost tracking, and observability. Provider-neutral, with no direct model access outside it.
Tenant isolation at three layers
Isolation runs at three layers: data, a per-tenant database with row-level security; access, scoped role-based retrieval; and infrastructure, VPC and network isolation. No shared data path between firms.
Hardened infrastructure
Secrets management, encrypted object storage and backups, monitoring and observability, and automated CI/CD. A Postgres primary, an isolated vector store, and full audit logs and history.
Private, single-tenant deployment
Firms that need it can run Bloxii single-tenant, in your preferred region, with customer-controlled data residency. The same platform on dedicated infrastructure.
Certifications & roadmap
Live, in progress, and on the roadmap.
GDPR
Built into the product from day one. UK / EU data residency, append-only audit, and customer rights honoured.
FCA-aware deployment
Deployed with FCA-aligned controls, MiFID II and AIFMD-aware reporting, human approval gates on every regulated output.
Cyber Essentials Plus
The UK government and FCA-aligned procurement baseline.
SOC 1 & SOC 2 Type II
Both SOC 1 and SOC 2 Type II, with evidence collected from day one. The control framework enterprise buyers expect.
ISO/IEC 42001
The AI Management System standard. Pulled forward so institutional buyers can procure with it as a table-stakes requirement.
Data handling principles
Your data stays your data.
Numbers are data, never guessed
Every figure Bloxii surfaces is pulled directly from your source records, Xero, Sage, fund admin, spreadsheets. The AI never estimates, rounds, or invents a number. If the data isn’t there, the output says so rather than filling the gap.
Every output cites its transaction
Regulatory and financial outputs carry a citation back to the exact source transaction or document. Your reviewer clicks through to the underlying record before approving, no claim stands without its evidence.
Human sign-off on every output
The AI does the operational work; your team reviews and approves. Nothing is submitted, filed, or sent without explicit human approval. You keep judgment and sign-off on every output, every time.
Adversarial cross-tenant isolation
Each firm runs in an isolated environment with no shared data path between tenants. The isolation extends to the retrieval layer: Ask Bloxii can never surface a document a user isn’t permissioned to see, access is enforced at retrieval, not just at display.
No training on your data, contractual and technical
Your data never trains shared models. Prompts and outputs are not used to train any model, enforced contractually in Bloxii’s Platform Agreement and with every model provider, and technically through the controlled AI gateway. Auditable on request.
UK residency · append-only 7-year audit
Your application and stored data are hosted in AWS London. Model processing depends on the models enabled for each firm and can be restricted to UK-hosted models where you require it. Every skill run, document touched, and human approval is written to an append-only log retained for seven years, inspectable by your auditors and regulators, exportable on demand.
Platform & architecture
How the platform stays trustworthy.
Grounded reasoning, not free-form generation
Bloxii builds Bloxii. Reasoning runs on a provider-neutral model layer, Claude on AWS London by default and the Bloxii model when it is ready, but every skill is constrained to your source data. The model reasons over your records; it does not generate financial figures from its own training.
Router across providers
Bloxii’s Router directs each task to the most suitable model your firm allows while holding your data-residency and no-training commitments as contractual obligations independent of any single provider. A material provider change triggers customer notification, provider independence is engineered in, not a hidden dependency.
Scoped, audited MCP access
Every skill and every connection call through Bloxii’s MCP endpoints is authenticated, scoped to specific capabilities, logged for audit, and bound by the same per-firm isolation as the rest of the platform. A skill has no broader reach than the user it acts for.
Role-based access control
SAML SSO, role-based access control, IP allow-listing, and data lifecycle management are standard for every firm from day one, not gated behind an enterprise plan. Access is enforced at the data layer, including retrieval.
Regression eval harness
Every skill runs against a regression eval harness before a change goes live. Skills are tested against known-good outputs so a tuning change can’t silently degrade accuracy on the work your firm depends on.
Enforceable, auditable commitments
Bloxii’s Security Addendum carries binding terms on data protection, access, no-training, and incident-response SLAs, aligned with SOC 2, ISO 42001, and GDPR. Independent third-party security testing is planned, external validation, not self-attestation.
AI data handling
The questions sophisticated buyers ask.
Private capital firms evaluating AI vendors ask AI-specific questions that traditional compliance frameworks don't fully cover, about accuracy, isolation, and who has the final say. Here's how Bloxii answers them, plainly.
Q01
Where does our data go when a skill runs?
Your data is stored in the UK. When a skill runs, the prompt goes from Bloxii’s UK infrastructure to the models your firm has allowed, through Bloxii’s controlled gateway, and the output comes back. Model processing is set per firm: keep your tenant entirely UK-only, or enable the specific models a team needs. Your data is never combined with another firm’s.
Q02
Is our data ever used to train AI models?
No, enforced two ways. Contractually: Bloxii’s agreements, both with you and with every model provider, prohibit training on your data. Technically: prompts and outputs are not used to train any model, and pass through the controlled AI gateway rather than a provider’s consumer service. The no-training guarantee is auditable on request.
Q03
Can the AI invent a number, or surface a document we’re not allowed to see?
No on both counts. Numbers are data, never guessed: every figure is queried from your source records and carries a citation back to the transaction behind it. And retrieval is permission-enforced: Ask Bloxii can never return a document a user isn’t entitled to access.
Q04
How are mistakes caught before they reach a regulator?
Regulatory and financial outputs cite their source data, and every output is reviewed and approved by a person on your team before it leaves the building. Behind the scenes, each skill runs against a regression eval harness so a tuning change can’t silently degrade accuracy.
Q05
Can we audit which model touched which document?
Yes. Every skill run is written to an append-only log, model, timestamp, input source, output, and approval state, retained for seven years. The full trail is inspectable in Bloxii's admin interface and exportable for audit or regulatory review.
Q06
What if a model provider changes its terms?
Bloxii’s Router runs across providers, Claude on AWS London by default and the Bloxii model when it is ready, so a single provider isn’t a single point of failure. Your residency and no-training commitments are obligations Bloxii holds independently of any provider. A material change triggers customer notification and migration if needed.