Application example. The scenario below was put together from what ProdMan builds in real projects, without representing a specific client. The figures shown are project estimates and vary with the number of sources, the volume of records and the rules of each business.
The scenario
In a mid-sized city in the countryside, looking for a property for sale means opening the website of every local agency, one by one. Each with its own filters, its own layout and, quite often, no price in the listing, just "contact us for pricing". Comparing two houses in different neighborhoods becomes an exercise in open tabs and a manual spreadsheet.
Whoever is searching does not want a system. They want an answer: is there a good house at this price, in this neighborhood, with this floor area?
The same pattern shows up in any operation with information scattered around: supplier quotes across five portals, a job posting published on several sites, public tenders, auctions, a competitor's price list. The subject changes, the problem is the same.
Data first, interface second
The first part of the work is not visual. It is collection. Every source has its own page structure, its own way of hiding or showing price, its own way of paginating results. Some require simulated navigation so the collection is not blocked. Others only reveal a value when the search is done by price range instead of an open listing.
- Multi-source collection: each source gets its own collector, adapted to the behavior of that site, including the ones that try to block automated collection.
- Deduplication: the same item listed by more than one source is identified and unified, with no repetition in the result list.
- Data preservation: when a source fails or comes back incomplete on a given day, the system keeps the last good result instead of showing an empty or broken list.
- Automatic refresh: a daily routine re-collects, deduplicates and publishes with no manual intervention.
The friction was in the clickable filter
The first panel already solves the aggregation: one single place, with filters for price, bedrooms, bathrooms, parking spaces, area, type and neighborhood. It works. Except a clickable filter repeats the same mental model as any portal: click, adjust, click again.
The question guiding the next stage is simple. Why does a person have to translate what they want into form fields, when they already know how to say it in one sentence?
They type the way they think and the system interprets it, extracts type, price, bedrooms and neighborhood, and applies the filters on its own:
- "house up to R$ 100k" becomes type house and maximum price of R$ 100,000.
- "two-bedroom apartment downtown" becomes type apartment, 2 bedrooms and neighborhood downtown.
- "lot over 300 square meters" becomes type lot and minimum area of 300 m².
The clickable filter stays on screen. It becomes the visible result of the interpretation instead of the starting point.
How it works
Typed sentence → AI interpretation → extraction of type, price, bedrooms and neighborhood → structured filters → aggregated base → sorted results.
The text search leans on the structured base. The result stays trustworthy even when the sentence is ambiguous: with no certainty about what was asked, the system applies the most likely filter and shows which filters were used, so the person corrects it with one click instead of rewriting the whole sentence.
How we thought about the risks
Interpreting a sentence is only useful when you can verify what was interpreted. The design accounts for that from day one.
- Every interpretation appears as a visible filter on screen, never as a silent result.
- What filters the data is the same structured logic as the manual mode. The AI handles intent, not the result.
- One-click correction: if the interpretation is wrong, the person removes or swaps a filter without retyping the sentence.
- No fabricated data. Price, neighborhood and floor area always come from the real collection, and the AI never fills a gap with an estimate.
- A source that changes its layout breaks collection silently, so the routine monitors volume drops and raises an alert instead of publishing a base full of holes.
- One single place to search, instead of five or more websites open in parallel
- The most common search resolved with no filter clicks at all, straight from the sentence
- Base refreshed automatically every day, with no manual operation
- Real comparison between items from different sources, with price and measurements in the same standard
- Figures are project estimates, calibrated during the diagnostic against the real sources of each operation
What the company needs to have
- The sources mapped: which websites, systems or spreadsheets hold the information today.
- Clarity about which fields matter for the decision, even if nobody can cross-reference them today.
- Authorization to use the data, respecting the terms of each source.
How this becomes a project
It starts with the free 45-minute diagnostic, where we map the sources, measure volume and check what can be collected. The build takes 1 to 4 weeks, as a Short Project: collectors per source, deduplication, unified base, panel, sentence search and the daily refresh routine. Delivered with documentation, a 90-day warranty and a client team that knows how to run it.