Assistant answering from your data
Structure and user flows that fit your need.
Free your team’s time
for valuable work.
We build assistants that work with your company knowledge, document processing and predictive models: retrieval over your own data, fine-tuning an open-source model on your data and running it on your own servers, forecasting and anomaly models; with data preparation and labelling before, human-approved automation after.
Needs analysis · Design · Development
YOUR NEED
“I want an assistant that works with my company’s knowledge.”
THE SOLUTION WE BUILD
From your knowledge to action
01Connect data
02Generate answers
03Approve & apply
A sample flow shaped around your project
WHAT DO YOU WANT TO DO?Your need is
Your need is
a starting point.
THE BUILDING BLOCKS OF YOUR IDEAWhat do
What do
we build?
We define the scope according to your use case and design each part as a piece of a working whole.
Chatbot & support flows
Development compatible with how you already work.
Fine-tuning & on-prem deployment
Testing and verification under real usage conditions.
Forecasting & analysis models
Delivery, documentation and continuation plan.
WHERE WE STARTNot a model —
Not a model —
a solved job.
An AI project does not start with "which model should we use". First we talk about which task eats time, who does it today and how, and what a correct answer looks like. Model choice comes after that, and usually the winner is not the most impressive model but the method that finishes the job cheapest and most predictably.
In practice we use four methods. Answering from your own data (RAG): your documents and records are made searchable, so the model does not guess an answer, it produces one from a source it found. Fine-tuning: an open-source model is trained on your data and learns your company's language, form structure and labels. Forecasting and analysis models: classic machine learning for numeric questions such as demand, stock, delay and anomaly — a large language model is often not needed at all here. Data preparation and labelling: the prerequisite for the other three; nothing works before scattered data is collected and cleaned.
Where the model runs is answered up front too. It can run against a cloud provider's API, or an open-source model can be installed on your own server so data never leaves your company. The two differ in cost, speed and privacy outcome; we decide which fits you in the first conversation.
CM Apps runs this setup in its own products too: we build and operate the assistants and automations in our own operation. In other words we carry the maintenance load of the setup we recommend before our customers do.
AI states a wrong answer just as confidently as a right one. That is why every flow we build shows the source an answer rests on, and operations with consequences are tied to human approval. We do not put a flow we cannot measure into production.
THE SHAPE OF THE FLOW
Your dataDocuments · database · email · panel records
Model layerRetrieval · fine-tuned model · forecasting model · rules
Your teamIn the panel, on WhatsApp or inside your existing software
The model is built on top of your data, it does not replace it. The source stays with you; the layer only reads, produces and sends for approval when needed.
HOW WE WORKSetup in
Setup in
four steps.
- 01
We understand the job and the data
How long does each operation take, where does each piece of data live, what does a correct answer look like? The output of this step is not a model but a measurable definition of success.
- 02
We choose the method
Is the answer inside a document, or is it a numeric forecast? Accordingly we pick RAG, fine-tuning, a classic model or a plain rule engine. Using an unnecessarily large model only inflates cost.
- 03
We build and connect
Data pipeline, model and interface are built; the result is connected to your panel, your existing software or your WhatsApp flow. Permissions and the audit log are defined in this step.
- 04
We measure and improve
It is tested with real questions, wrong answers are collected, and the data or the model is corrected accordingly. Improvement is continuous, not a one-off.
WHAT WE BUILDMethods and
Methods and
what we deliver.
- Assistant answering from your data (RAG)Your documents are chunked and made searchable; the assistant produces answers with a cited source instead of inventing them.
- Model fine-tuningAn open-source model is trained on your company data and learns your terminology, forms and labels.
- Deployment on your own serverThe model runs on your company's server; data does not leave. We plan the hardware need together up front.
- Forecasting & analysis modelsDemand, stock, delay and price forecasting; anomaly and fault detection. Classic machine learning for numeric questions.
- Document & data processingField extraction from invoices, contracts and delivery notes; turning free text into structured records.
- Data preparation & labellingCollecting, cleaning and labelling scattered data and building the pipeline. The prerequisite for everything else.
- Chatbot & support flowsA support assistant on your website, panel and WhatsApp; when it cannot answer, it hands the conversation to a human.
- Human-approved automationOperations with consequences (payment, contract, deletion) never run alone; the approval step is written into the flow.
HOW IT IS USEDScenarios
Scenarios
from the field.
"What was this customer's contract term last year?"
The contract archive becomes searchable; the assistant finds the clause and tells you which article of which document it is.
Who keys in the incoming invoice?
Fields are extracted from the document and the accounting entry is prepared as a draft; staff review instead of typing.
How much of which product will we need next month?
Demand forecasting is built from past sales, season and campaign data; an order suggestion lands in the panel.
Support answers the same question forty times a day
Frequent questions are answered from company knowledge; exceptions go to a human, so the team only handles the hard ones.
We notice production faults too late
Anomaly detection is set up on sensor and production records; an alert fires when the deviation threshold is crossed.
Our data cannot leave the building
An open-source model is installed on your server and fine-tuned on your data; no record goes to the internet.
DATA AND BOUNDARIESWhat goes where
What goes where
is explicit.
- Where data is processed, in writingCloud API or your own server — the choice is made up front and written into the contract. It does not change quietly later.
- Training opt-out by defaultYour data not being used to train the provider's model is the default; the opposite only happens at your explicit request.
- Field-level maskingIdentity, contact and financial fields are marked up front so they never reach the model.
- Sources are mandatoryEvery answer the assistant produces states which document it rests on; an answer that cannot be verified is not an answer.
- Approval-bound operationsIrreversible operations such as money, contracts and deletion do not run without human approval.
- Audit logWhich question was asked, which source was read, which operation ran — all of it is recorded.
QUESTIONS ON YOUR MIND
Before we start.
Will our data be used to train the AI provider's model?+
By default, no. When a cloud API is used we run with the training opt-out; with an open-source model installed on your own server the data never leaves your company anyway. Which setup was chosen is written into the contract.
Is fine-tuning or RAG right for us?+
If the answer is inside your documents and the content changes often, RAG is more accurate and cheaper. If the model needs to learn your language, forms and labels, fine-tuning comes in. They are also used together; we decide by looking at the data.
How much data is needed?+
It depends on the method: RAG works with your existing documents and needs no extra collection. Fine-tuning and forecasting models need historical records; we look at what you have and tell you plainly whether it is enough or whether data preparation comes first.
What do we need to run it on our own server?+
A GPU or a strong server, sized to the model. Small models can run on existing hardware; large ones need separate investment. After scoping we give you the required hardware and its alternatives in writing.
What happens if it gives a wrong answer?+
The source behind an answer is always shown, so errors become visible. Operations with consequences are tied to human approval. Wrong answers are collected and the data or model is corrected — this is a permanent part of the work.
Do we have to replace our existing software?+
Usually not. The AI layer is built on top of your existing system and reads data through an API or the database. The result is connected back to your panel, your existing software or your WhatsApp flow.
Do the source code and the model stay with us?+
Yes. The code we develop, the data pipeline and the fine-tuned model weights are delivered to you. The setup keeps working even if you do not continue with us.
How long until we see results?+
It depends on scope: an assistant built for a narrow set of questions goes live quickly, while multi-source or fine-tuned setups need data preparation. In the first conversation we narrow the scope and lay out a realistic order.
IN OUR FIRST MEETINGFirst, let us
First, let us
listen to you.
You can come with an idea or an existing system. Let’s start with these questions and map out an actionable roadmap.
Let’s plan together- 01
Which task takes the most time?
- 02
Which data sources can be used?
- 03
How will we evaluate the answers?
PART OF A BIGGER PROJECT?
Areas of expertise that work together.
NEXT STEPTell us about your business.
Let’s talk about your project