IntentHub
A multi-tenant B2B platform that finds the companies in the market to buy, right now. It reads messy, unstructured web data and turns it into clean, scored purchase-intent signals on a real-time dashboard, so sales teams reach out the moment intent appears instead of guessing. Built for a France-based client and live in production.
Buying intent, buried in noise
Sales teams sit on a mountain of public web data about their accounts. Almost none of it is structured, and most of it is noise.
Finding the few companies actually in-market means hours of manual digging, and by the time someone notices, the moment has already passed.
What if the noise scored itself, and only the accounts worth a call ever surfaced?
The brief
A France-based client needed to spot high-intent prospects from messy web data without a human reviewing every row.
The bar was reliability at scale: turn noise into a signal a sales team would actually trust, cheaply enough to run on every account, every day.
Noise in, signal out
A pipeline that reads the open web and hands a sales team a short, ranked list of accounts worth their time.
A live intent feed
Signals stream onto a real-time dashboard, scored and ranked, so the hottest accounts surface the moment they start heating up instead of showing up in a report next week.
Scored by Claude
Every signal runs through Claude with a strict JSON schema. Scoring is contextual: it reads the pattern across a company's activity and writes a plain-language rationale, not a weighted sum of clicks.
Signal agents
Configurable agents watch for the signals that matter to each team, with ICP filters, per-signal weights, and hot and warm thresholds that decide when an account is worth a call.
Multi-source intelligence
Seven signal streams across first, second, and third-party sources, from a website tracking pixel to funding news and hiring data, unified into one score per account.
How it works
An async pipeline that scrapes the web, extracts with an LLM, and only emits the signals it is actually confident about.
Inside the product
Command center
The live intent feed: scored accounts, ranked hot leads, and the signals behind each one.
Account intelligence
Per-account AI insight and full intent history.
Signal agents
Churn defense
Integrations
Landing page
Tech stack
Async task orchestration, real-time WebSockets, and a trustworthy LLM layer, all multi-tenant.
Backend
- Django REST Framework
- Celery
- Redis
- PostgreSQL
- Daphne / Channels
Frontend
- Next.js
- React
- TypeScript
- Tailwind CSS
- TanStack Query
- Zustand
AI layer
- Claude API
- Strict JSON schema
- Confidence + abstention
- Ground-truth evals
Infra + data
- Docker
- Multi-tenant
- WebSockets
- n8n automations
- HubSpot / Apollo
Decisions that mattered
The calls that shaped the build, and the reasoning behind each one.
Async over inline
Scraping and LLM calls are slow and bursty, so they never run in the request path. Celery and Redis split the work into decoupled stages: scrape, analyze, extract, publish. One stage failing does not cascade, and the expensive stages scale on their own.
Abstention over hallucination
A wrong signal costs a client a wasted sales call, so the system is allowed to say I don't know. Low confidence or thin evidence holds a signal back instead of inventing intent. That came straight out of my MSc dissertation on faithfulness and abstention.
Evals over vibes
A small hand-labeled ground-truth set measures extraction precision, so prompts, how much context to feed the model, and the abstention threshold are all set by numbers, not gut feel.
Structured output, always
Claude is constrained to a strict JSON schema, so every extraction is validated data the rest of the pipeline can trust, not a paragraph to parse downstream.
Multi-tenant from day one
Every record is organization-scoped with role-based access, so it is a real SaaS with isolation between clients, not a single-tenant script with a login bolted on.
Real-time by default
Daphne and WebSockets push new signals onto the dashboard as they are scored, so the feed is genuinely live instead of a page a rep has to keep refreshing.
What's next
Immediate roadmap
- Full public launch from limited release
- Bring your own key for Claude to cut per-account cost
- Deeper enrichment caching
- More third-party signal sources
Future vision
- Self-serve onboarding
- Per-tenant scoring that sharpens with feedback
- Deeper CRM and outreach automation
- SOC 2 and enterprise controls
Need this kind of AI pipeline?
IntentHub is Claude turning raw web data into structured, trustworthy signals, in production for a real client. If that is the shape of what you are building, let's talk.