Two candidates for one venue:
· "Café du Coin"
· "Cafe du Coin (Ville)" — 40m away
Same place, or two?
Messy sources in, data you can build on out. An AI-assisted pipeline verifies, deduplicates, and ranks every record — so what you get back is actually trustworthy.
No scraping scripts to babysit, no RevOps team, no pile of half-cleaned CSVs. Point Zonzent at a source and describe the shape you need — the AI-assisted pipeline enriches, verifies, and ranks the rest.
Zonzent isn't a demo. Lokozy — a curated city discovery app — is the first product built entirely on the pipeline. Every venue it shows was enriched, validated, and ranked by Zonzent.
A discovery app that helps you find the right place to eat, drink, and go — starting with Luxembourg City. Its entire venue dataset comes out of the Zonzent pipeline: cleaned, deduplicated, geo-verified, and scored.
Lokozy is client #1 — one instance of general-purpose infrastructure, not a bespoke tool. The next city is just a new config file.
Visit lokozy.com ↗Each record moves through a fixed pipeline. What changes between projects — city, entity type, output — lives in configuration, not code.
Pulls raw entities from the source — OpenStreetMap, Places, or a CSV you hand over.
Fills the gaps — hours, contact, attributes — and cross-checks each field against reality.
Verifies each record against authoritative place data, bounded by a hard geographic rectangle.
Ranks, structures, and writes clean records to your store — deduped and scored.
Clean data isn't about pulling more rows. It's about knowing which rows are lying to you. A few decisions that keep the output honest:
A venue can match a name 100% and still be 188km away. Matching is validated by geography — a hard bounding box plus a 1km rejection threshold — never by name similarity alone.
Permanently and temporarily closed records are excluded before they ever reach an end user. Reopenings self-heal on the next run. 0 closed venues shipped.
Open data carries junk — dead listings, wrong categories, out-of-area noise. The pipeline assumes this and filters for it rather than trusting the feed.
Some decisions can't be made by a regex. Is "Café du Coin" the same place as "Cafe du Coin (Ville)"? Is a bar tagged restaurant actually a restaurant? Is this listing closed, or just missing hours?
Zonzent puts a language model exactly where that judgment is needed — resolving entities, disambiguating categories, reading status — then hands the verdict back to the pipeline to validate against hard geographic and freshness checks.
Reasoning: Claude · bounded by geo + status rules · never the last word alone.
Anywhere the source is scattered, stale, or half-trustworthy and you need clean structured records to build on.
Venues, points of interest, local businesses — enriched, geo-verified, ranked. The dataset behind a discovery app or map.
Live — powering LokozyA rough list of companies in, structured briefs out — profiles, signals, and context, without a RevOps team to run it.
Near-termBring a source that's scattered or stale. Get back deduplicated, verified, confidence-scored records — cleaned to a schema you define.
Let's scope itCity, entity type, bounding box, output collection — all configuration. The pipeline logic never changes. That's what makes the same infrastructure serve restaurants in one city and a completely different dataset in the next.
Point it somewhere new, set the bounds, run it.
The pipeline is general-purpose. Live capability first — everything below is the honest state, not a wishlist.
The engine running in production today.
New config, new bounding box, same pipeline.
CSV in, enriched company briefs out — research without a RevOps team.