Discovery is a research project
The right API, file, or table is rarely the first search result. Often it's an undocumented portal API that takes days to find.
Zillusion · agentic web-data platform
The web is the world's largest database. We make it queryable.
Xscent Global Inc. builds Zillusion. Describe the data you need in one sentence; agents find the sources, build and verify the scrapers, and keep the datasets fresh. Every result passes an independent verification gate first.
you ▸Global CO₂ emissions by country — CSV or JSON
discoverfan-out: 4 search engines · vertical registries · 23k-row API catalog (semantic)
discover13 sources found → scored on 5 dimensions · grouped APIs / files / embedded
top hitGovernment of Canada — Greenhouse Gas Emissions (CESI) · score 8.6
exploreprobing access paths (API? CSV? page table?) → writing workflow.py
validateisolated validator re-runs workflow.py in a fresh session…
validateresampled rows re-fetched → values match
validatefield meanings re-proven against the live page
gatescorecard complete → verdict: PASS
runfull crawl: 63 records · 3 CSV files · exit 0 · 9.5s clean
warehousedataset v1 + manifest saved · monitor armed → diff & notify
Web data is a maintenance treadmill.
Anyone who needs web data pays the same three taxes.
Most of the cost is discovery and maintenance — not analysis.
Scrapers rot
Site redesigns silently break selectors. The worst failures aren't crashes — a selector wired to the wrong field keeps returning plausible, wrong data.
“Looks right” isn't validation
Spot checks don't scale, and a model reviewing its own scraper approves it. Nobody re-checks the data against the live page.
One sentence in. A monitored dataset out.
The whole lifecycle runs as one conversation.
Discover
Searches 4 engines, vertical registries, and a 23,000-row API catalog with bilingual semantic retrieval. Candidates are scored on 5 dimensions and grouped as APIs, files, or embedded data.
dense embeddings + BM25 · reciprocal rank fusionBuild & validate
Agents probe the live site and write workflow.py, preferring APIs over brittle selectors. An isolated validator re-runs it; a gate computes the verdict.
terminal routes: deterministic · agentic · inline · infeasibleWarehouse & monitor
Datasets are versioned into a warehouse with incremental diffs. Schedules re-run robots; drift and failures notify you by webhook or email.
versioned datasets · scheduled re-runs · diff + notifyEngineered so it cannot lie to you.
Four bets: one hard to copy, three that deepen with use.
The interface can be copied. The verified datasets and solved walls underneath can't.
isolated · adversarialVerification a model can't sweet-talk
A separate read-only agent re-runs every scraper in a fresh session; it cannot edit what it grades. Most tools let the same agent write and approve the scraper.
gate-computedThe verdict is computed, not argued
PASS or FAIL comes from a scorecard: re-runs clean, reproducible, resample matches, field meanings re-checked against the live page. Not from a model saying it looks good.
discover + steerIt finds the source, you keep the wheel
One sentence in, and it ranks every public path to the data: API, file, or page. Most tools need you to supply the URL. At a login or captcha wall, you take over in an embedded browser, then hand back.
compounds with useEvery solved job makes the next one cheaper
Access playbooks and per-site extractors persist and become cross-site skills; a shared data layer makes the hundredth request for a dataset far cheaper than the first. The skill reuse is live today.
An honest map of the landscape.
A big, crowded, fast-moving market. Who does what, and where we differ.
~$25BData-as-a-Service market · 2025Mordor Intelligence
≈17%/yrAI web-scraping segment growthFuture Market Insights, 2026
33% by 2028of enterprise apps will use agentic AI (from <1% in 2024)Gartner, 2025
Research & coding agentsPerplexity · ChatGPT · Claude · Manus
Answer a question or produce a one-off file. None runs a repeatable, monitored data pipeline.
where they winGeneral reasoning, zero setup.
No-code scrapersBrowse AI · Octoparse · Apify
You point at a site and configure each robot; page changes mean re-tuning.
where they winMature point-and-click, huge template libraries.
Crawl & extract APIsFirecrawl · Zyte · Diffbot
Fetch or auto-extract from a URL you supply. Choosing the source, verifying, and hosting stay your job.
where they winManaged proxies, JS rendering, scale.
Data infrastructureBright Data & co.
Sell proxies and pre-collected dataset catalogs. You still pick what to buy or configure.
where they winIndustrial unblocking, compliance, delivery SLAs.
Zillusionby Xscent Global Inc.
From one sentence: find and rank sources, build the extractor, verify it against the live site, warehouse and monitor. Steerable at every step.
what's differentDiscovery + isolated verification + monitoring, in one loop.
Kadoa, Nimble, and Reworkd are racing at the same loop. Our bet: plain-language source discovery across every public path, and a PASS/FAIL computed by an isolated gate rather than claimed by the model. Firecrawl and SearXNG run as components inside Zillusion.
Where we are.
The platform runs end-to-end in production. What's left is productization, not research.
Shipped
- End-to-end pipeline live: discover → build → validate → run → warehouse → monitor
- Conversational orchestrator with a per-user robot library
- 23k-row API catalog, bilingual semantic retrieval
- Anti-bot memory reused across sites
- Bilingual UI (EN / 中文), deployed on cloud infra
Now
- Hardening against live sites at scale
- Usage-metered billing on prepaid credits (design finalized)
- Early design-partner conversations
Next
- Hosted multi-tenant offering
- Data-product layer: audited cleaning, reports, derived datasets
- Managed proxy pool
23,000+ rows in the curated API catalog5 scoring dimensions per source4 gate checks before a PASS500+ backend tests green
Xscent Global Inc.
We're building the query layer for the open web. It's early, and we're talking to investors who back infrastructure.
or write directly: baibaze35@gmail.com