Case file04
Open models run on site
Legal and IT management had ruled out the public cloud. Assisted document search exists nonetheless, running on machines hosted on the premises.
- LOCAL AI INFRASTRUCTURE
- FRANCE / MONACO / SWITZERLAND
- REFERENCE ON REQUEST
What the client entrusted to us
An organisation held a large volume of sensitive documentation — files, contracts, minutes and technical notes accumulated over decades. Finding precise information meant knowing where to look, and often asking the person who had handled the file.
A public-cloud assistant experiment had been stopped by legal and IT management: the very principle of sending documents to a third-party service was incompatible with commitments made to clients.
The request was therefore twofold: obtain genuinely useful document search, and demonstrate that nothing leaves the perimeter. We handled both together, because the second condition determines the first one’s architecture.
The starting point
What the diagnosis revealed.
Before any technical proposal, we write down the real state — including what unsettles. This description, validated by the teams, is what makes the rest verifiable.
- 01
People-dependent search
Information often sat with whoever had handled the file, not in an index everyone could consult.
- 02
Heterogeneous documents
Multiple formats, old scans, parallel folder trees, with no shared repository and no naming rules.
- 03
Cloud assistant ruled out
Sending documents to a third-party service had been refused by legal and IT management.
- 04
No traceability of answers
An assistant that answers without quoting its source cannot serve in a binding professional setting.
- 05
Unpredictable costs
Usage-based billing made any budget projection difficult and dependent on query volume.
- 06
No in-house skills
Nobody on site knew how to operate or maintain an artificial intelligence model in production.
Constraints
What could not be compromised.
- 01
No data leaves the perimeter
No call to an external interface, no telemetry, no training on the organisation’s documents.
- 02
Answers must quote their source
Every result had to point to the document and passage it came from, with no exception.
- 03
A human validates
The tool assists search; it neither decides nor produces anything authoritative by itself.
- 04
A predictable cost
Investment had to be hardware-based and controlled, not proportional to request counts.
- 05
Operable on site
In-house teams had to restart, monitor and update the whole without us.
- 06
Reversibility
Models and indexes had to remain replaceable components, with no single-vendor dependency.
The AIGYROS answer
What we built.
We started with the material: corpus inventory, format analysis, access-perimeter definition. A document is only queryable if we know who may read it — this rule structured everything else.
The chosen architecture is local: dedicated machines hosted on the premises, open models run on site, an in-house document index and a search layer returning source passages.
Answers are presented with references: document, section, passage. The user opens the source, checks and decides. No output is presented as a conclusion.
Operation was documented and transferred: restart procedures, hardware monitoring, model updates, index management. In-house teams run the whole; we remain in support.
Conduct
Four stages, held in this order.
Stated durations are orders of conduct magnitude, never commitments: each engagement is recalibrated on its own ground.
- 01Corpus analysis phase
Frame
Document and format inventory, access-perimeter definition, priority use cases chosen with teams.
- Document and format inventory
- Access-perimeter definition
- Priority use cases with teams
- Written expected-quality criteria
- 02Architecture phase
Size
Hardware and open-model choice, local architecture and network isolation, indexing strategy.
- Hardware and open-model choice
- Local architecture and network isolation
- Indexing and update strategy
- Operating-cost estimate
- 03Deployment phase
Install
Machine and model installation, authorised-corpus indexing, search interface with quotations.
- Machine and model installation
- Authorised-corpus indexing
- Search interface with quotations
- Quality and no-leak tests
- 04Operations phase
Transfer
Restart and monitoring procedures, model updates, in-house team training.
- Restart and monitoring procedures
- Model and index updates
- In-house team training
- Periodic usage review
What changes
Before, after — no showcase figures.
We do not publish improvement percentages: they cannot be verified from outside. What can be described, on the other hand, is the daily gesture that changes.
- Initial situationAfter the engagement
Initial situation: Information located with people
After the engagement: Indexed corpus, queryable in natural language
Initial situation: Cloud assistant refused by legal
After the engagement: On-site models, no transfer
Initial situation: Sourceless answers
After the engagement: Every result quotes document and passage
Initial situation: Usage-proportional costs
After the engagement: Hardware investment, predictable operating cost
Initial situation: Single-vendor dependency
After the engagement: Open models and reversible indexes
Initial situation: No in-house skills
After the engagement: Documented operation transferred to teams
Deliverables
What stays in your hands.
An engagement is also judged by what it leaves behind. Every deliverable is written to be read without us.
- 01
Corpus inventory
The documents concerned, their formats and access rules, validated by management.
- 02
Local architecture
Machines, network isolation and continuity plan for the dedicated environment.
- 03
Document index
The indexed corpus, with its update and removal procedure.
- 04
Search interface
Natural-language search, with mandatory quotations and source opening.
- 05
Operating procedures
Restart, monitoring, model updates, incident handling.
- 06
Test report
Result quality and verification of no external transfer.
Excluded scope
What the engagement did not include.
A blurry scope costs more than an incomplete one. So we write it down in black and white, from the proposal stage: what is handled, and what is not.
- Supplying or rewriting documents: corpora belong to the organisation.
- Any automated decision: the tool assists search, it does not decide.
- Using public-cloud artificial intelligence services, ruled out by the legal framing.
- Training models on proprietary data, unnecessary for the stated need.
What never leavescannot leak.
Case file
Open models run on site
Case 04 — Local AI infrastructure
Area of operation: France / Monaco / Switzerland
Other case files
Prendre contact
Does your situation resemble one of these?
Tell us about your environment, your constraints and what cannot be compromised. We answer with analysis, not a catalogue.