F-01
Private model deploy & training
On-prem LLM + fine-tune
Open-source models run on company servers and can be trained or fine-tuned on internal data; conversations and data stay inside the network.
← Back home
CASE FILE / HM-AI-002
Enterprise self-hosted AI
Open-source models on your own servers, trained or fine-tuned on internal data; documents indexed as a RAG knowledge base and systems wired up as MCP tools — data never leaves your network.
01 / PROBLEMChallenge
Staff want AI; IT cannot approve public clouds
Public AI services raise data-residency and confidentiality concerns. Internal knowledge is scattered across documents, manuals, and systems — one question means searching several places, and answers are not always trustworthy.
02 / SOLUTIONSolution
Models in-house, answers with sources
A private LLM built on open-source models runs on company servers, and we train or fine-tune it on company data when needed. Internal documents are indexed so questions are answered by retrieval with citations, and internal systems connect as MCP tools the assistant can query directly.
03 / FEATURESFeatures
Asked in-house, answered with evidence
F-01 Private model deploy & training
F-02 RAG knowledge base
F-03 Custom MCP tools
F-04 Permissions and audit
How many annual leave days?
Per the Staff Handbook §3.2, annual leave is 12 days, plus up to 3 days by seniority.
Staff Handbook §3.2
HR Policy 2026
How are stock variances handled?
Variances above 2% require a discrepancy ticket, verified by the warehouse manager and closed in ERP.
Stock Policy v4
Warehouse Manual §7
Where is the Q3 sales report?
The latest one is in “Report library › Sales › 2026-Q3”, or ask the BI group for a copy.
Report library index
BI group notice
Internal docs → index→Ask → retrieve + MCP → cited answer
F-02
RAG knowledge base
Document index + retrieval
Manuals, policies, and internal docs are indexed; answers cite their sources and link back.
F-03
Custom MCP tools
Internal systems as tools
ERP, ticketing, and CRM systems connect as tools the assistant can query and act on.
F-04
Permissions and audit
Roles + audit trail
Visibility is scoped by role; conversations and tool calls are logged for audit.
04 / OUTCOMEOutcome
Qualitative outcomes
- O-01Ask in-house, answers with sources
Staff get grounded answers directly instead of hunting through systems.
- O-02Data never leaves the network
Models, indexes, and conversations all live on company servers.
- O-03Internal systems become AI tools
With MCP wiring, the assistant checks stock and tickets on your behalf.
Next: start with your system.
If this is close to the problem you need solved, get in touch.
Go to contact