Build a vertical on top of WebRobot
A vertical — sentiment analysis, lead generation, price intelligence, a trading desk — solves one problem end to end. You don’t rebuild the engine each time: the platform is the tier, and a vertical is a thin application plus a plugin that reuses the same distributed stack.
Describe what you need — the tier does it
Your vertical calls WebRobot for the capabilities that are expensive to build and operate. It holds none of them.
Web acquisition
A real-browser fleet with anti-bot resilience, proxies and consent handling.
Distributed processing
Apache Spark on Kubernetes, with a live catalogue of pipeline stages.
AI agents
Multi-agent teams on Ray that design, validate and run work, with human approval on critical steps.
Governance & BYOC
Hosted or your own LLMs; orgs, roles, API keys and audit; run it all on the customer’s cloud.
A curated source catalog behind every vertical
610 sources across 34 markets — news, forums, reviews, e-commerce, social, bookmakers and more — each already classified by how to fetch it (simple HTTP vs a real browser), its anti-bot posture and whether it is reachable now. Every vertical — sentiment, price intelligence, lead generation, a trading desk — draws on it, so an agent picks the right source and method without rediscovering them. Sources behind strong anti-bot are scoped case by case, and platforms are accessed within their terms of use.
One bundle, four component types
A vertical extends the platform through a single <plugin>-<version>.zip. Each component is loaded by the layer it belongs to, validated as a whole, enabled per organisation, and — with hot-reload — published without restarting the process.
ETL stages
Scala/Spark stages that run inside the distributed engine and appear in the live catalogue the moment they’re installed — a new verb for every pipeline.
REST API
Jersey endpoints serving chart-ready queries and agent-callable tools, scoped to the caller’s organisation.
CLI
Subcommands added to the WebRobot CLI, discovered by the host — the same capability from the terminal.
Dashboard UI
Next.js views hot-loaded into the console as ESM — a vertical’s own screens, inside the platform.
The application layer bends the stack — without forking it
The distributed runtime (Spark, Ray, Kubernetes) is fixed and operated once. The plugin layer is how a vertical shapes it to a domain.
New capability, same infra
A stage runs on the existing Spark cluster; an agent profile on the existing Ray cluster. You add a verb, not a cluster.
Discoverable by agents
Because the catalogue is live, the agentic pipeline-designer sees your new stages immediately and composes them — no prompt surgery, no redeploy.
Per-tenant, not per-fork
Enablement is scoped to an organisation; a vertical rolls out to the tenants that need it, each with its own credentials, data and billing.
Runs where the data is
The same bundle installs on the shared platform or a BYOC cluster on the customer’s own cloud — the app moves, the tier follows.
How to build one well
Keep the app thin
No scraping and no ML in your app — those belong to the tier. The app describes the goal and shows the result.
Let an agent design the pipeline
Launch a background agent on Ray that generates and validates the pipeline against the live catalogue — not a single LLM call.
Validate before you run
Dry-run the manifest (shape) and the pipeline (sample rows). A pipeline that validates but collects nothing is not done.
Compose from the catalogue
Pick stage names from what the catalogue returns, never from memory — tenant plugins add to it.
Scope every key
Call the API with an organisation-scoped key from your server, never a platform-admin key, and never from the browser.
Ship a bundle
Package the components with a manifest and install once. Version it; enable it per org; let hot-reload publish it without a restart.
Sentiment analysis
The user describes what they want to understand in plain words. A background agent designs a sentiment pipeline — acquire → extract → analyse → persist — from the live catalogue, validates it, and runs it on the platform. Rich sentiment (polarity, a Plutchik emotion vector, named entities and aspect-based scoring) is produced by an ETL stage shipped as a plugin, and the results come back through the vertical’s own REST endpoints and dashboard views. The engine is WebRobot; the vertical is the brief and the insight.
GoCyprus
AI travel guide and lead generator for Cyprus
Pipelines keep events, places and local products fresh; an AI guide plans trips in conversation and books directly with local businesses (hotels, gyms, tours) through an MCP catalog of their offers. Each booking is a qualified lead for the business.
acquire events → enrich → MCP catalog → guide agent → booking
Open the preview (under construction) →GeneaAI
AI genealogy assistant for Italian families
Records published by comuni, parishes and archives are indexed for search; the assistant builds the family tree as you tell it, and every finding comes with its source — you decide what enters the tree. The publishers earn a share whenever their documents are used.
archive records → index → search agent → sourced findings → family tree
Open the preview (under construction) →GainOnDelta
Real-estate arbitrage for Italy and the UK
Per-country pipelines collect for-sale listings and score each against comparables in the same area and official data (HM Land Registry in the UK). An AI analyst discards false bargains, checks the area on the web and writes a printable report with shortlist, benchmarks, risks and sources.
listings → comparables + official data → score → analyst agent → report
Open the preview →