KAPLAKATANext demo
Meridian FreightSAPPHIRE · AI TEAMMATETry Sapphire

Case study · Meridian Freight Co. · Doral, FL

An AI teammate that knows your business — and never leaks it.

Sapphire lives in Meridian's company chat. She reads the ops handbook, queries the TMS, drafts carrier emails and asks a person before anything leaves the building. The model, the documents and the logs all run inside Meridian's own private cloud.

Try the live demo ↓See how it's built
  • Self-hosted
  • Your data stays in your VPC
  • Human-in-the-loop
  • Audit log
kickoff to rollout
5 days
documents sent to an outside AI service
0
of outbound actions approved by a person
100%
median answer time in pilot
1.8 s

Live demo

Watch Sapphire work.

Pick a scenario or type your own question. The trace shows every tool call with its arguments, latency and token cost, plus the audit log your security team receives.

demo data

#ops

Sapphire online · in-VPC

Dana RuizDispatch lead

Dock 3 at the Doral yard is clear again. Thanks for the quick turn, everyone.

SapphireAIAI teammate

Morning, ops. I'm connected to the ops handbook, the TMS and carrier email. I only see what you can see, and I'll always ask before sending anything outside Meridian. Ask me something or pick a prompt below.

Scripted demo: replies and tool results are canned so it runs anywhere. Production builds call a model hosted in the client's own cloud.

Architecture

Everything runs inside the fence.

Chat messages reach an agent runtime in Meridian's private cloud. Every tool call is checked against policy on the way out, and every step lands in the audit log.

Sapphire architectureMessages from company chat go to the agent runtime inside Meridian's private cloud. Every tool call passes through a guardrails and policy layer before reaching the knowledge base, TMS and ERP, email, or CRM. Every step is written to an append-only audit log.MERIDIAN PRIVATE CLOUD · VPCCompany chat#ops · DMsAgent runtimeopen-weight modelAudit logappend-onlyGuardrails · policy · approvalsKnowledge baseRAG over SOPs & contractsTMS / ERPloads · lanes · invoicesEmailsends only on approvalCRMaccounts · read-scoped

  • Retrieval, not training

    Docs are indexed in a private vector store and quoted at answer time, so updates show up the same day.

  • Permission-aware

    Each tool call carries the asking user's identity. Sapphire can never see more than they can.

  • Swap the model, keep the system

    The runtime is model-agnostic. Upgrade to a newer open-weight model without rewiring tools.

Beyond freight

Same pattern, other teams.

The parts that make Sapphire safe are not specific to logistics. A few possibilities for other industries:

  • Legal research

    Could answer “what did we agree on indemnity with this vendor?” from your own contract archive, with the exact clause cited.

  • Finance close

    Could reconcile sub-ledgers, flag variances over a threshold and draft journal entries for a controller to approve.

  • Property management

    Could triage maintenance requests, look up lease terms and draft owner notices, with a manager signing off before anything is sent.

  • Healthcare intake

    Could pre-fill intake forms and route questions to the right staff while patient data stays in your environment. Clinical decisions stay with clinicians.

How it's delivered

Live in about five days.

One focused week, one person accountable, and a working assistant in your team's chat at the end of it.

  1. Day 1

    Discovery

    Pick the three to five jobs the assistant should own, which docs it may read, which tools it may touch, and who approves what.

  2. Day 2–3

    Connect docs & tools

    Index handbooks and contracts for retrieval, connect read APIs for the TMS or ERP, email drafts and chat. All inside your cloud account.

  3. Day 4

    Guardrails & testing

    Write the policies as code, try to break the refusals, and test against real questions collected from your team.

  4. Day 5

    Rollout + training

    Launch in one channel, a 45-minute training session for the team, and a runbook for your admins.

Security

Private by architecture, not by promise.

  • Runs in your cloud

    Model, vector index and logs live in your own account or on-prem servers. No calls to a third-party AI API.

  • Nothing is trained on your data

    Sapphire retrieves from your docs at question time. Nothing is used to train a shared model.

  • Inherits permissions

    Sapphire only sees what the person asking can see. A dispatcher can't read finance through her.

  • Approval gates

    Every write action, like sending email or changing a record, waits for a named approver.

  • Policy as code

    Data rules such as “no bulk exports” are versioned config, reviewed like any other code change.

  • Immutable audit trail

    Every prompt, tool call and decision is logged append-only and can stream to your SIEM.

guardrails/meridian.yamlexample
# no call ever leaves the VPC
model:
  host: in-vpc
  retention: none
access:
  inherit_user_permissions: true
tools:
  send_email:  { approval: required }
  update_tms:  { scope: notes_only }
  crm.export:  { allowed: false, rule: DLP-03 }
audit:
  sink: siem://meridian-sec
  immutable: true

Want a Sapphire for your team?

A private assistant on your infrastructure, connected only to what you choose, with a person approving every action. Typically live in about five days.

Start a project with Kaplakata →Run the demo again