On-Premises Personal Agents
Run your own personal-agent platform on your own servers, giving each employee or customer a personalized agent that remembers them and acts for them within the permissions you set.
Solutions
Know what your AI did—and what the records can actually support
AI Governance & Evidence
Our Forward Deployed Engineers (FDEs) build and run an evidence store on your own infrastructure. The engagement brings scoped event intake, agent and model inventories, and case-review work alongside FinSafe, FinShield and FinGuard records and policy enforcement. Named reviewers reconstruct what happened, separate observations from assumptions and keep evidence gaps visible.
- Isolation and policy boundaries around execution.
- Rule checks and a hash-chained action audit.
- An FDE-built evidence store on your infrastructure, with human-led review.
Ways to start
Build around one workflow
choose its owner, records and business-system routes. Our FDEs build the store and intake for that scope, and configure an agent and model inventory with your team. People confirm ownership and update responsibilities.
Investigate one case
begin with the question to answer, not a request to collect everything. Assign an analyst and agree who may access sensitive content. Use existing records before deciding whether additional collection is necessary.
Review
When someone asks what happened
Distinguish the attempted action, control decision and downstream result. A policy entry alone does not prove a business-system change; an application result alone does not explain the boundary. Review the connections, and state where the available records cannot establish them.
Solutions
A case and its reconstruction
Case side
a team reports that an assistant attempted the same business update twice. The case owner records the reported time, the intended change and the affected workflow. The initial report is a question to investigate, not a finding.
Analyst side
the reviewer checks available execution records, the relevant FinGuard journal entries and the business system's result. Where request identifiers and timing support a connection, the reviewer records the basis. Where they do not, the gap remains explicit. A duplicate response in the gateway record is assessed alongside the downstream result; no missing event is invented to complete the story.
Solutions
Contain, decide, review
Contain
use FinSafe isolation or desktop FinShield to constrain the execution environment. Check the applicable permissions and available records; do not infer that isolation makes the generated content correct.
Decide
FinGuard applies configured rules to actions routed through it. Its idempotency ledger handles repeat writes, credential custody keeps enterprise secrets outside agents, and its journal and hash-chained audit record the action decisions.
Review
our FDEs build and run the store and case-handling setup; people assess the available material and write findings. Keep source references with each conclusion and name the follow-up owner. Our FDEs configure human approval for selected FinGuard action paths and tamper-evident evidence bundles for case review. Reviewers still assess coverage and missing records.
Privacy
Privacy and accountability
Request only what the case needs. Limit access to full content, separate working notes from original records and agree handling and retention with the owners. Where possible, share an aggregate summary rather than raw personal material. The organisation decides what to report and to whom; a vendor or assistant does not take that responsibility.
Products
How it is built
- FinSafeisolation, runtime tracing and associated policy records. Its OS boundary is distinct from the meaning of an agent’s answer or tool request.
- FinShielddesktop boundaries and policy records. On Windows, supported conversations, tool calls and results can also be retained locally according to configuration. Local records are not a promise of central content collection.
- FinGuardthe stateful, rule-based action gateway with credential custody, duplicate handling, journal and hash-chained audit. It is not a replacement for model or tool gateways.
FDE-built evidence store
our engineers build and operate the store on your infrastructure, configure telemetry and OTLP intake for agreed sources, and maintain the agent/model inventory setup with your team. This service layer is distinct from the products and their own records.
FDE
Work with FDE
- Scope: agree records, owners and infrastructure.
- Build: our FDEs implement the store, source intake and inventory setup, with the selected gateway approval and evidence-bundle workflow.
- Check: reconstruct a sample case with named reviewers and document gaps.
- Run: our FDEs operate the agreed setup with your team; document access, handling, retention choices and responsibilities.
FAQ
Common questions
Can a sandbox stop hallucination?
No. It constrains execution and access, not the truth of an answer.
Can a privileged host administrator bypass containment?
Yes. Host privileges are a boundary of the protection and must be considered in the review.
Does FinGuard cover every action automatically?
No. Coverage depends on routing actions through it, controlling accepted credentials and constraining network paths. Examine bypass routes.
Does local collection mean the model runs locally?
No. Model processing is separate. Check the endpoint and where prompts and tool output go.
Who builds and runs the evidence store?
Our FDEs build and operate it on your infrastructure as an engagement, not a self-serve software package. Your team owns the findings and reporting decisions. Available records and configured intake define the scope; a store does not make missing evidence complete.
Next step
Bring one workflow to our FDEs.
We build and run its evidence-store setup on your infrastructure, while your team keeps responsibility for review, gaps and conclusions.