AI content engine and lead‑capture website
A company website with a content engine built in: it gathers industry news, scores each story, and drafts posts in the firm’s voice for a person to approve. Its forms turn each inquiry into a CRM lead, with spam screened out first.
- By September 2026 the Maple DX blog had 15 published posts, each approved by a person first.
What it does
The firm’s blog is built to react to its industry’s news with its own take, and nobody starts a post from a blank page: Claude writes the drafts, and a person decides what goes live. Each inquiry from the site goes into the CRM as a lead, with an email alert as a backup and spam screened out first.
The situation
Maple DX, a Salesforce consulting firm, sells to the technology leaders of mid-sized companies. It wanted a blog that answers each week’s news about the AI and data platforms those buyers use, in the voice of a senior practitioner, while the news is still fresh. Writing every post from scratch is slow work.
The site’s forms also had to put each inquiry into the firm’s CRM, and an open web form draws spam that clutters a CRM and can send confirmation emails to strangers.
What was built
A website with a content pipeline built in. On a schedule it gathers recent stories from a list of industry news feeds and web searches, then opens each article to confirm its real publish date and screen out event teasers with no article behind them. Claude, Anthropic’s AI model, scores each story against a written rubric, groups stories that cover the same news, and drafts posts in the firm’s voice from the strongest ones.
Every draft waits in a private review queue. A person edits it as needed and publishes it or sets it aside; the pipeline itself never publishes. Each week Claude also proposes tutorial topics based on the news it has scored, and a person approves each topic before it is drafted.
The site’s forms send each inquiry to the firm’s CRM as a new lead, with the person’s answers and the source they came from, and email an alert to the team at the same time. Spam is screened out before either happens, and the person’s confirmation email goes out only once the inquiry is safely recorded.
How it works
Gather the news
Every six hours the system reads a list of industry news feeds and web searches. Stories more than a week old, or with no publish date, are dropped, and a link that arrives twice is stored once.
Check each article
Before any AI is used, the system opens each article and reads the page’s own publish date, because feeds often restamp old stories. Event and webinar teasers with no real article behind them are set aside.
Score for fit
Every hour Claude scores new stories from 0 to 100 on five measures: relevance to the firm’s buyers, how fresh the insight is, how authoritative the source is, whether it gives the firm a strong angle, and how likely a buyer is to read it. It also suggests the call to action the post should end with.
Group and draft
Stories that score 70 or more are grouped when they cover the same news, so several reports on one launch become one post. Claude drafts each post from the articles themselves, under written rules: name a product only when sure who makes it, keep to a set length, and avoid phrases that read as machine-written.
A person approves
Drafts wait in a private review queue. A person can edit the title, text, call to action, and tags, then publish or archive each draft, and a published post can be pulled back at any time. News posts link to their original source, and each post gets its own preview image for sharing.
Screen out spam
The site’s forms post to the firm’s own server, which drops requests sent from other websites, forms where a hidden field that people never see has been filled in, and forms finished faster than a person could. A dropped request still gets an ordinary success reply, so a bot learns nothing about what stopped it.
Record the lead
Each inquiry becomes a lead in the CRM with the person’s answers and the source they came from, and an alert email goes to the team at the same moment. If either one succeeds, the inquiry is kept, and the person’s confirmation email goes out only after it is safely recorded.
How it is tested
The pipeline was switched on one stage at a time, on real sources. Scoring ran first, with nothing drafted, until a person had checked its judgments. That check caught a feed that had restamped an old story’s date and an event teaser that scored well, which led to the date and teaser checks that run before scoring. Drafting came next, and the publish button was added only after a person had reviewed the first drafts. That review found a product credited to the wrong company, two drafts from the same article, and drafts that ran short, and each led to a fix in the drafter.
Every stage can also be run on demand through a private admin trigger, which was used to check each one on live sources. On the lead endpoint, every spam check was tested by sending a request it should drop and confirming both the ordinary-looking reply and the server’s log entry, and a real submission was tested through the site in a browser.
Tools, status, and who built it
- Tools
TypeScript, Next.js, React, Tailwind CSS, Supabase (Postgres), Netlify, Anthropic Claude, Exa, Salesforce, Resend
- Status
Live
- Who built it
Stratyx Ventures
Talk to our team about your project
Whatever you’re building or fixing, you’ll leave the first call with a straight answer on what it would take.
After the call, AI work starts with a Workflow Review for $2,500, and a site or an app with a written scope you approve.