An AI that works alongside the team
Avisa Aí uses a multimodal vision-language model, running on its own infrastructure and orchestrated by nine specialized routines. It reads photos, videos and text to filter out what is invalid, flag risks and save the team’s time.
Principles
The AI was designed to be useful without taking the decisions that matter away from people.
When in doubt, a person decides
When the model’s confidence is low, no automatic decision is recorded: the case stays in the human moderation queue.
The same rules as a moderator
The AI has no shortcuts: every decision it makes goes through the same permissions and server validations that apply to the team.
Photos stay in-house
The model runs on its own server with dedicated processing. Citizens’ images are never sent to commercial AI services.
Suggests, never rewrites
Suggested category, priority, title and description are visible to the maintenance team, but the report’s original fields are never changed without a person.
The nine routines
Each routine has a single responsibility and is triggered by a system event or on a schedule.
- A
Step moderation
On every action recorded on a reportChecks whether the attached evidence (approve, start, complete, validate) matches the problem. Clearly invalid evidence on a positive action reverses the step.
- B
Overdue deadlines
PeriodicallyFinds reports stuck past their step deadline and proposes the decision the responsible person would make, always revalidated by the server.
- C
Validation on creation
As soon as the report is sentClassifies what the photo or video shows, corrects the category when needed, flags image quality problems and rejects offensive text.
- D
Summaries for management
Once a dayWrites an executive summary of the indicators and the moderation queue, shown on the city dashboard.
- E
Enrichment
When the report is approvedSuggests the best category, a risk priority from 1 to 5, a clearer title and description, and flags sensitive data visible in the image. Nothing is applied without the team.
- F
Content report triage
When someone flags contentPre-classifies the flag and records a suggestion for the moderator, who makes the final call.
- G
Feedback triage
When a user sends feedbackOrganizes suggestions and criticism with tags, sentiment and a one-line summary.
- H
Natural-language search
When a manager searches the dashboardTurns a question like “urgent potholes from last week” into the matching filters on the report list.
- I
Comment moderation
When a comment is postedCatches insults the automatic dictionary does not know and replaces only the offensive part with ***.
Details that make a difference
Technical decisions drawn from real use of the system.
Video as evidence
Videos are analyzed by the AI too, and a frame from the video becomes the report’s cover.
Risk that changes the deadline
When the AI assigns top priority (for example, a “pothole” whose photo shows a burst pipe), the response deadline drops to the shortest in the system, whatever the category.
Two layers against offensive text
A hand-reviewed dictionary censors swear words and insults before the text is saved; the AI covers what the dictionary does not know, such as other languages and disguised insults.
Never blocks the citizen
Analysis runs in the background after the report is saved. If the AI service is down, the report proceeds normally and is reviewed by people.
Security and privacy
How Avisa Aí protects accounts, limits abuse and handles citizen and city data.