ERP
Orders, quotes and finance in plain language
Ask about open orders, late deliveries or a customer's history; get reports and summaries written for management.
Masked: customers, suppliers, amounts
Masked AI gateway
Claude for your company. Your data stays inside.
Lemy.Gate sits in your server room between your people, your systems and Claude. Names, codes and prices are replaced inside the box before anything leaves; the answer is restored inside the box before anyone reads it.
Lemy.Gate
On-premise appliance with a local masking model, your data dictionary and one controlled route to Claude.
Illustrative rendering
Today
Your teams already paste company data into AI.
Quotes, customer e-mails, cost sheets and drawings go into public AI tools every day. Banning them rarely works; the data simply leaves without a record.
The Lemy.Gate approach
The strongest models, without handing over the raw data.
Every request passes through Lemy.Gate. Sensitive values are found and replaced on your own hardware; Claude only sees placeholders. Requests that must not leave at all are answered by the local model.
Scope of service
- InstallationLemy.Gate in your server room: hardware, local model, network and the single outbound route.
- Data definitionsYour customers, suppliers, staff, parts and prices become the masking dictionary, each class with its own policy.
- OperationClaude-powered assistants and agents across ERP, MRP, MES, WMS, PDM and CRM, with every outbound request logged.
Server room
Right next to your existing systems.
Lemy.Gate is installed in the same server room as your ERP, MES, WMS and PDM servers. No changes to existing systems are required.
Data definitions
It learns your vocabulary from your own master data.
Read-only connectors pull customer, supplier, employee, part and price lists. They become the dictionary the gateway masks with: exact matches, not guesses.
Company-wide
Every department, one controlled exit.
Production, warehouse, engineering, sales and finance use the same gateway. Every AI request takes the same route, under the same rules.
- ERP · MRP
- MES · WMS
- PDM · CAD
- CRM · Office
Controlled exit
Only placeholders cross the boundary.
Names, part numbers and amounts are replaced inside the box. Claude works on the structure of the question; the real values are put back inside the box.
Claude's intelligence. Your data stays yours.
Illustrative rendering · Blender
Use the best AI
without giving away
your data.
Local models keep data inside but fall short on reasoning, writing and long tasks. Cloud models are strong but see everything you send. Lemy.Gate splits the work: your hardware handles the data, Claude handles the language and the reasoning.
Layered masking
Your data dictionary first, then fixed patterns (tax ID, IBAN, phone, e-mail), then a local model for what is left. Placeholders stay consistent across a conversation.
Predefined data classes
Each class of data has a policy you set: mask, generalise, keep local or allow. Some classes never leave, whatever the request.
Numbers stay local
Totals, margins and comparisons are calculated in the box. Claude receives a template with placeholders and writes the text; the box fills in the figures.
Every outbound word logged
The audit log shows exactly what left, when and for whom. Users can preview the masked text before it is sent.
How it works
Mask, ask,
restore.
One request, end to end. The mapping between placeholders and real values is stored encrypted in the box and never leaves it.
01 · INSIDE · USER
The request
Write a reply to Ahmet Yılmaz at ABC Makina: the 4.2 M TRY quote for part TK-4410 can ship on 14 Nov.
Typed in the assistant, in WhatsApp, or triggered by an ERP event.
02 · OUTSIDE · CLAUDE
What Claude sees
Write a reply to [PERSON_1] at [CUSTOMER_1]: the [AMOUNT_1] quote for part [PART_1] can ship on [DATE_1].
Only this text crosses the boundary, over one allow-listed route, and is written to the audit log.
03 · INSIDE · USER
The answer
Dear Mr Yılmaz, we're pleased to confirm that part TK-4410 under our 4.2 M TRY quote for ABC Makina can ship on 14 Nov…
Claude's draft comes back with placeholders; the box restores the real values before the user reads it.
Example with invented names and values.
Predefined data definitions
You decide
what may leave.
Data classes are defined once, during installation, from your own systems. The policies below are a starting point; each company sets its own.
| Data class | Source | Default policy | What Claude receives |
|---|---|---|---|
| Customers, suppliers, contacts | ERP · CRM | Mask | [CUSTOMER_1], [SUPPLIER_2], [PERSON_3] |
| Part numbers, drawings, revisions | PDM · ERP | Mask | [PART_1], [DRAWING_1] with type and revision kept |
| Prices, costs, margins | ERP · MRP | Keep local | Placeholders only; figures are calculated and filled in inside the box |
| Quantities, stock, lead times | MRP · WMS · MES | Generalise | Ranges and relative values ("about 2 weeks late", "+18%") |
| Personal IDs, IBAN, phone, e-mail | All sources | Mask | [ID_1], [IBAN_1], [PHONE_1] |
| Payroll, health, recipes, defence projects | Defined per company | Keep local | Nothing; answered by the local model or refused |
| General questions, public information | — | Allow | The request as written |
Systems
A smarter layer
over the systems
you already run.
Lemy.Gate connects to each system through named, read-only functions. Claude works on the masked result: it explains, drafts, plans and flags, while the data and the figures stay with you.
MRP
Material shortages explained before they bite
Shortage lists turned into prioritised purchasing actions, with the reasoning written out for the planner.
Masked: parts, suppliers · local: quantities, costs
MES
Shop-floor status without opening a screen
Station progress, stoppages and delays summarised per shift; operators can report and ask over messaging.
Masked: work orders, operators, customers
WMS
Stock and shipments, answered directly
What is ready to ship, what is blocked, which locations need counting, in one answer instead of three reports.
Masked: customers, parts · generalised: quantities
PDM
Engineering knowledge you can search
Find the right revision, compare change notes and draft engineering change summaries.
Masked: part and drawing numbers · local: drawings themselves
CRM
Customer follow-up that writes itself
Meeting notes logged from free text, follow-up e-mails drafted, quotes and reminders tracked.
Masked: people, companies, amounts
Free assistant
Claude for every employee, through the same gate.
Everyday writing, translation, analysis and presentations, the way people already use AI, but through one controlled route instead of personal accounts. Users see what will leave before it is sent.
Masked: every class in your data definitions · logged: every outbound request
Principle
Your data.
Your dictionary.
Your keys.
The appliance sits in your server room, the mapping between placeholders and real values is encrypted inside it, and the keys and access rights stay with you. Lemy.co installs, defines and operates; it doesn't look inside.
Trust & compliance
What we
claim, and what
we don't.
Masking reduces what leaves the company; it does not make the data disappear. We would rather say that plainly than overpromise.
Masked is not anonymous
Because the mapping stays with you, masked data can still count as personal data under KVKK and GDPR. Each deployment starts with a legal review of cross-border transfer.
Detection is measured, not assumed
No masking method catches everything. During the pilot we measure detection on your own documents and report what was missed.
Some data never leaves
Classes marked "keep local" are never sent, whatever the request. For organisations that cannot send anything at all, the appliance also runs fully closed.
One route, fully logged
The appliance has one outbound route, allow-listed to the AI provider's API. Everything sent over it is recorded and reviewable.
Add-on services
Extend it to
fit your needs.
Add any of these services to the base installation individually. Scope and pricing for each are agreed with you based on your processes.
Lemy.Act
Agents that carry out defined operations
Claude proposes; the box executes predefined operations such as purchase requests, work orders and stock transfers. Every action is previewed, committed only with authorised approval and written to the audit log.
- Defined action catalogue only
- Single or dual approval by risk level
- Action limits and undo steps
Lemy.Routine
Scheduled routine operations
Morning reports, delay follow-ups, purchasing reminders and permission reviews run on schedule, with the same masking rules.
- Daily and weekly summaries
- Delay and shortage alerts
- Permission and dictionary reviews
Lemy.Connect
Messaging channels
Use the assistant from WhatsApp, Telegram, Teams or internal e-mail. Messages pass through the same gateway, so the same data definitions apply.
- Internal channels: e-mail, corporate messaging
- External channels: WhatsApp, Telegram, Teams
- Role-based answers per user
When external messaging channels are enabled, messages are relayed through the respective provider's infrastructure; your organisation decides which data classes may be shared over each channel.
Hardware tiers
Sized to
your scale.
We don't have customer figures yet; these are the technical specifications of the hardware classes we propose. Because Claude does the heavy reasoning, the box mainly needs to detect, mask and route. The final configuration is set during the pilot.
128GB
Gate · unified memory
NVIDIA DGX Spark class. Masking, routing and a local model for everyday local answers.
96GB VRAM
Gate Plus · per card
RTX PRO 6000 Blackwell class. Adds a 70B-class local model for the data classes that must stay inside.
1
Outbound route
Allow-listed to the AI provider's API only. No other traffic leaves the appliance.
0
Mapping tables sent out
The link between placeholders and real values is encrypted and stays in the box.
Hardware figures are manufacturer specifications. NVIDIA, Claude and Anthropic are trademarks of their owners and are named only to identify the product class and the model provider; Lemy.co is not affiliated with them.
Pilot programme
Pilot
partnership.
We have no published case studies yet. With our first pilot partners we define the data classes on their own systems and measure masking on their own documents, prioritising machining and machine manufacturers running ERP with MES or WMS.