Data leaves with every question
Every query to an external AI sends a contract, report or ticket outside. Without an answer to that, the project stops at the DPO.
Why a model of your own
The AI pilot works. Then compliance asks where the data goes. And finance asks what it costs at real volume.
In regulated industries, sending customer, patient or operational data to a third-party server stalls the project in legal. And usage-based billing grows along with the AI's success.
The bill grows with success
Usage-based billing rises with volume. When the AI works and scales, the cost comes in higher than planned, and the price changes whenever the vendor wants.
Showcase benchmarks, not your case
The vendor shows its model winning a public test. Nobody shows whether it gets your contract, your report, your ticket right.
What you get
A decision backed by proof, and a model under your control.
The gain does not come from the biggest model. It comes from choosing the right one for the task, proving it on your cases and running it under the same security rules the rest of your company already follows.
01
Sensitive data never leaves the house
Data preparation, tuning and model responses all happen in your data center or private cloud. A team with certified information security and privacy runs the process.
ISO/IEC 27701certified privacy
02
Decisions with results in hand
Before investing in tuning, we compare small models and the API on your real cases, with the estimated cost of each path. If the API wins, the recommendation is the API.
03
Human review where mistakes are costly
Every stage has a named approver, and the model's critical decisions in production go through a person. It is the same method as our deliveries with agents.
97%right the first time
04
Costs you can predict
A small model running on your infrastructure has a cost per task that does not rise with every question. You plan the budget without depending on someone else's price list.
05
No owner but you
The tuned model, prepared data, case set and playbook stay with you. When a better model comes out, you switch based on the test, not the promise.
How we work
Five stages. A human gate in each one.
Each stage produces the input for the next, and nothing moves forward without approval from whoever answers for the result in your business. The data never leaves your environment at any stage.
Task and success criteria
A specific task, the cost of a mistake and the yardstick that says whether the AI got it right, written before any test.
Gate · Acceptance criteria signed by the businessData in your environment
Data organized, cleaned and anonymized where LGPD (Brazil's data protection law) requires it, with access rules and recorded lineage.
Gate · DPO or compliance clearanceReal case set
Situations from your operations, with the right answer known, to test any model the same way.
Gate · Cases validated by specialistsTuning and comparison
The tuned small model and the API answer the same cases. We compare accuracy, response time and cost.
Gate · Decision: own model, API or neitherProduction and operations
The model in your environment, with quality monitoring, human review of critical decisions and a reassessment with every new version.
Gate · Periodic performance review
What changes in your AI operations
DataStops traveling. It stays in your environment, with controlled and logged access.
CostNo longer tracks the number of questions and becomes an infrastructure cost you can plan.
AuditHas an answer: where the data is, how the model was tuned and how it is tested.
Model changesBecome your decision, with the case set telling you whether the new model is better.
MistakesStop being a surprise. A drop in quality triggers an alert before it reaches the customer.
How the market sells it
Everyone promises AI under control. Few prove it on your case.
Those who sell small models usually own the model or the platform, and prove it with public tests. We have no model to sell, so the test is on your cases.
How to engage
Three formats. All of them start with proof.
You can stop after any of them. Each format delivers a decision or a model that stays with you.
Start here
Feasibility assessment
Who it is for: For those with a candidate task who need to decide between their own model and the API.
3 to 4 week format
- One task chosen with the business and written success criteria
- A real case set validated by your specialists
- Comparison of small models and the API on the same cases
- Estimated running cost for both paths
- A written recommendation, including when it is to use the API
Most requested
Single-task pilot
Who it is for: For those who already have the task defined and want to see the tuned model running in their own environment.
8 to 12 week format
- Data preparation and anonymization in your environment
- Model tuning for the task
- Side by side evaluation with the API on the case set
- Controlled use with real users and human review
- Compliance dossier: data, access, tuning and tests
Ongoing operations
SLM in production
Who it is for: For those putting the model into a process who need it stable, auditable and up to date.
Ongoing contract, with quarterly review
- Deployment in your data center or private cloud
- Monitoring of quality, usage and cost
- Reassessment with every new model version
- Retuning with new data when operations change
- Access control and audit trail
Scope and price come out of the conversation about your task and, later, the feasibility assessment. The timelines above are the format we propose, not a promise of results.
26 years of critical software, with information security and privacy certified by external audit.
Goes well with
From process to agent, with your data protected.
Frequently asked questions
Before you start.
What is an SLM?
Small Language Model: a small language model that understands and writes text like the better-known ones, only smaller and specialized. Because it is smaller, it runs on the company's own servers, costs less to operate and can be tuned to the vocabulary and rules of your business.
Is a small model good enough?
For a well-defined task, a tuned small model usually comes close to the large ones, and Gartner predicts that by 2027 companies will use small, task-specific models three times more than large general-purpose ones (Gartner, 2025). For open, varied tasks, the large model may be better. That is why we test both with your cases before recommending.
When do you recommend not using an SLM?
When the data is not sensitive, the volume is low or the task requires broad knowledge. In those cases, a large model via API is usually simpler and cheaper, and that is what the recommendation says.
Where does the model run and what infrastructure do I need?
In your data center (on-premise) or your private cloud. Inference happens in your environment. Small models, with quantization, usually run on one or a few GPUs, and the feasibility assessment already sizes the hardware.
How does this help with LGPD and my industry's rules?
The data stays in your environment, with access control and an audit trail, and we deliver a dossier on where the data is, who accesses it, how the model was tuned and how it is tested. That is the kind of answer regulators such as BACEN (Brazil's central bank), ANS and ANEEL usually ask for. DB1 is ISO/IEC 27001 and ISO/IEC 27701 certified.
Is the tuned model mine?
The tuning, the prepared data, the case set and the playbook stay with you. The base model follows the original vendor's license, which is why the license is a selection criterion from the very first stage.