06 Data, AI & platform
Applied AI & business automation
Language models and automation inside defined workflows — drafting, checking, extracting and routing — with a person approving what matters.
Outcome: Less manual work, with a person accountable for every decision.
Applied AI here means a large language model (LLM) working inside one defined step of an existing process: drafting a letter from case data, checking records for gaps, extracting fields from an uploaded document, routing a request. Each feature has a clear input, a structured result and a person who confirms it.
The best candidates are tasks where people spend hours re-typing, summarising or checking documents that follow a pattern, and where the result can be reviewed quickly. It does not fit decisions that need professional judgement, or tasks where nobody can tell a good answer from a bad one.
Typical needs
- Hours lost to drafting, summarising or re-typing documents
- Repetitive data entry between systems
- Interest in AI, but concern about where company data goes
What you receive
- AI features scoped to specific, reviewable tasks
- Model calls from the server, with data minimisation and usage limits
- Automation of the repetitive steps between your tools
Approach
How the work runs.
- 01
One task, one definition of done
A narrow task is chosen, and what a correct result looks like is written down with real examples from your data before any model is called.
- 02
Only the data the task needs
Context is loaded with the permissions of the person asking, personal identifiers are masked where the task allows, and models are called from the server, never from the browser.
- 03
Proposals a person confirms
The model returns its result in a fixed structure that the application checks, and it appears as a proposal. Nothing is written, deleted or sent until someone accepts it.
- 04
Limits, logs and consent
Each user has usage limits, a log records who used which feature without storing the request text, and AI is switched on only with the organisation’s consent, then per user.
Technologies
- OpenAI API
- Anthropic API
- Local models (Ollama)
- Python
- Structured outputs
The engineer who does this work is the founder, Filip Volf — you work with him directly.
Questions
About this service.
Where does our data go when a model is called?
To the model provider you choose, from the server, and only the data the task needs. In Pametni Advokat, for example, the assistant masks emails, phone numbers, JMBG and account numbers in common formats before text reaches the AI service, and response storage is turned off. Where data must stay on your own infrastructure, local models are an option.
Can the AI change records or send messages by itself?
No. Features are designed so the model prepares and a person decides. In Pametni Advokat, AI document import writes only the entries the lawyer confirms, and drafts are a starting point the lawyer edits.
How is this different from ML & AI R&D prototypes?
Applied AI puts a model behind a feature people use at work, with human approval and the controls above. ML & AI R&D prototypes answer open questions first: experiments in simulated environments or on offline data, with results reported together with their limits. A prototype that proves useful can later become an applied feature.
Related services
- 05
APIs, backend architecture & PostgreSQL
Outcome: Business rules enforced once, where the data lives.
- 07
ML & AI R&D prototypes
Outcome: Evidence to build, rework or stop, from experiments you can rerun.
- 09
Security-aware architecture
Outcome: Security decisions made deliberately — and written down.
Next step
Let’s talk about what you’re building.
Describe what you need built or fixed, how the work runs today and the constraints. You get clarifying questions or a first outline of the approach — by email, in English or Serbian.