RAR-AI

Results

steps from header to review, navigable in any order

11steps from header to review, navigable in any order

of field data (drafts, photos and audio) kept on the device, with no server

100%of field data (drafts, photos and audio) kept on the device, with no server

touch targets on the main controls of the custom design system

44 pxtouch targets on the main controls of the custom design system

Screens and components

1. 01

   Field collection

   Without signal, the indicator switches to offline and filling in continues, with each item rated in a single tap. Each vulnerability gets numbered photos, and the voice note is kept as audio, with partial transcription where the browser supports it.

   • 

     Enlarge

     
   • 

     Enlarge

     
   • 

     Enlarge

     

2. 02

   The same codebase on a computer

   On a computer, the same codebase adds a sidebar with each step's status: complete, in progress or not started.

   rar-ai · etapa 5

   

   Enlarge

   

3. 03

   Review and transfer

   The review lists what is missing in each step, with a shortcut to open the step and finish it. With no server, the report travels between devices as a JSON package, with or without photos and audio.

   • 

     Enlarge

     
   • 

     Enlarge

     

1. 01

   Offline saving on the device

   Pain

   Data is collected in places with an unstable connection. Any network dependency mid-inspection risks losing what has already been recorded.

   Solution

   Every change is written immediately: the report goes to localStorage and photos and audio go to IndexedDB, to stay clear of localStorage's ~5 MB limit. An indicator shows “Saved · time” and flags when the device is offline, without interrupting the work, and the report reopens where the analyst left off.

   Impact: Saved on every change, no save button

2. 02

   Confidential rules engine

   Pain

   The report depends on extensive domain logic — risk level, mandatory checklists and a conclusion consistent with the findings — which, outside the app, required a separate calculation, the analyst's memory and retyping.

   Solution

   A deterministic TypeScript rules engine computes the risk level automatically, builds the mandatory checklist based on the unit's profile and generates the conclusion draft from the recorded vulnerabilities and recommendations. The methodology and rules are confidential and are not detailed here.

   Impact: Risk calculation covered by automated tests

3. 03

   Steps with status and open items

   Pain

   Analysts walk the unit in the order the site dictates, not the template's, and in a report with dozens of fields it is easy to hand in something incomplete without noticing.

   Solution

   The template became 11 steps that open in any order, with a step bar showing each status and listing specific open items (e.g. “Vulnerability 02: description, impact and evidence”). The review gathers them all with a shortcut to each step and warns before the report is generated.

   Impact: Review is marked complete only when the other 10 steps are

4. 04

   Voice notes with transcription

   Pain

   There are many long text fields, and typing on a phone during the visit is slow.

   Solution

   Every long text field has a recorder: the audio is kept as a note, with playback and “Insert into text”. Where the browser offers the Web Speech API and there is a connection, speech is transcribed into the field in Brazilian Portuguese.

   Impact: Voice on every long text field, from the assessment to the conclusion

5. 05

   Numbered photo evidence

   Pain

   The template requires each vulnerability to point to its photo evidence, numbered in the report.

   Solution

   Photos come from the rear camera or the gallery, are compressed and numbered automatically (“Image 01”), and can be reused for another vulnerability. With no photo and no “evidence not applicable”, the vulnerability stays an open item.

6. 06

   Standardized assessment of the axes

   Pain

   The official template asks for a rating and observed situation for dozens of items; on a phone that means constant scrolling and lookups.

   Solution

   All axes follow the same pattern: observed situation per item and, for control items, a one-tap rating, with the assessment parameters at hand; then vulnerabilities recorded from the rated items themselves and recommendations picked from a predefined list.

7. 07

   Pre-drafted conclusion

   Pain

   The conclusion repeats vulnerabilities and measures already recorded in the axes, which forced analysts to retype the same content.

   Solution

   The conclusion fields that repeat what was recorded in the axes have a button that inserts an editable draft built by rules from that record. There is no AI involved: the text comes only from what the analyst recorded.

8. 08

   Official report as PDF

   Pain

   The final report combines tables, results and photo evidence that would otherwise come from separate notes and files.

   Solution

   The report page assembles the document following the official template's structure from what was filled in, with the rules engine's results and the photo evidence, plus a dedicated print style for Print → Save as PDF.

9. 09

   Transfer package (.json)

   Pain

   Analysts collect on the phone but review and finish on a computer, or hand the visit to a colleague. With no server, the draft would be stuck on the device where it started.

   Solution

   The report is exported as a JSON package, with or without embedded photos and audio, and imported on another device. Import is validated by the same Zod schema as the state and points to the incompatible field when a file doesn't fit.