Skip to content
Documentation
Overview
What FACE is

What FACE is

FACE runs multi-agent analysis over an organisation’s own operational data and turns it into decisions a person can review, approve and act on. It consists of a Go gRPC backend and a Flutter web application called the cockpit.

The core loop: fetch → extract → reason → recommend 

    flowchart LR
    A[Fetch<br/>connect & sample your systems] --> B[Extract<br/>cluster, chunk, ground]
    B --> C[Reason<br/>sovereign models + statistical tools]
    C --> D[Recommend<br/>action cards, evidence documents]
    D -->|approve / execute| E[Your systems & people]
  1. Fetch. FACE connects to your data sources through a configured connection, either directly or through a runner you control, and samples what it needs. See Data sources.
  2. Extract. Records are grouped into data domains, chunked, and indexed for retrieval, so later reasoning is grounded in your records rather than in a model’s general knowledge.
  3. Reason. Language and vision models served inside your cluster, plus classical statistics and machine learning (clustering, gradient-boosted trees, time-series decomposition, causal and Bayesian analysis), work over the extracted data. See Analysis & evidence.
  4. Recommend. Results arrive in the cockpit as action cards and evidence documents: specific, reviewable outputs such as an over/short/damage report, a temperature-excursion record, or a notice of intent to claim.

What it is used for

The product is built around operational problems in logistics, fulfilment and field operations:

  • Claims recovery. Cross-checking measured weights and counts against shipping documents, and drafting the resulting claim paperwork as evidence documents.
  • Cold chain and condition monitoring. Reading telemetry from temperature, humidity and shock sensors, then detecting and documenting excursions against stated thresholds. FACE never assumes a threshold the source didn’t provide.
  • Chain of custody and customs. Auditing custody records and flagging customs manifest exceptions.
  • Reverse logistics. Triaging returned items.
  • Compliance, finance and operations analysis. Assessments with a remediation workflow: findings are assigned, decided and routed.
  • Visual inspection. Grading frames from RTSP/ONVIF camera feeds with an in-cluster vision model, plus optional detection of individual objects of a named class (for example pallets or trucks). Detection is off by default.
  • Scenario analysis. Simulating “what if” questions against your data before committing to a change.

Principles you can rely on

  • Sovereign AI. Every model call goes to a model plane that runs in your cluster. FACE sends your data to no third-party LLM API.
  • Your data stays yours. Connection credentials are sealed and never returned by the API. Data at rest is encrypted, and FACE refuses to start without its encryption key. See Security & trust.
  • Honest outputs. “No data” and “could not look” are reported as different things, a connection test that didn’t actually reach the source says so, and documents FACE cannot support from evidence are deliberately not produced.
  • Recommendations wait for a person. Action cards stay pending in a queue until someone approves or rejects them. A remediation decision is stored with the authenticated identity of the person who made it, and FACE refuses to record a decision that has no author.
Architecture