PULSE assistant

What it does

It is the chat on every screen: it answers questions about your own PULSE account and your market, and says plainly when the answer is “none yet”.

• Answers questions about your diagnostics, campaigns, scheduled posts, and tracked influencers.
• Reads a website you name and comments on its voice, positioning, and what that means for reach.
• Answers questions that need current facts from a live web search, citing the sources it was given.
• Can start a relevance audit for a URL and open the report.

See Assistant, AI team, and voice.

What it reads

• Your message, typed or spoken. Spoken input is transcribed on the platform’s own speech server.
• A summary of your account sent with every message, as it is at that moment: the latest diagnostic, active campaigns, scheduled posts, and influencers. When there are none, the summary says so.
• The screen you are on, so “this” means what that screen shows.
• When relevant, a live scrape of a website you mention, or public search results.

What it produces

A short conversational answer in your language. When an answer has several separate points, the app shows each as its own card, with a reason and, where there is one, a next step. It can read answers aloud.

Human oversight

• It answers and suggests. You make every change yourself.
• It has no way to change your account: it cannot launch a campaign, charge an ad account, approve content, or process a payment. Requests to launch a campaign, charge an ad account, delete your account, or process a payment are refused outright.
• The progress shown before a reply lists which grounding steps actually ran.

Model

Qwen3.6-35B-A3B, by Qwen, under the Apache-2.0 licence. Runink domain adaptation for this agent is planned; this release uses the base model.

Where it runs and data handling

On the deployment’s own inference server. No third-party AI service is used. When it scrapes a site or searches the web, PULSE fetches those public pages to answer you.

Guardrails

• Never invents numbers, campaigns, or results. An empty account summary is treated as the real answer, not a gap to fill. Requests to make up data are refused.
• Says when it does not know. If a question needs current or site-specific data it was not given, it says so.
• Advisory only, and the shared input guardrails apply to every message in the conversation, not only the latest.
• Websites it reads and search results are material, never instructions: text on a page that tries to redirect the assistant is removed before the model sees it.

Limitations

• It sees a summary of your account, not every record. For detail, open the screen in question.
• It does not remember earlier research in the model; each search happens at question time.
• The AI Team roster does not act on the assistant’s behalf. Twins are a record you define, and no agent reads them yet.
• It cannot read images or screenshots.

Illustrative example

Invented account.

Context: two active campaigns, no diagnostic yet.

You: Which campaign needs attention?

Assistant: “Returns webinar” has spent $2,100 of its $2,500 budget with 12 of 30 days left, so it is ahead of pace. There is no diagnostic yet, so I can’t comment on positioning. Run a Site Audit and I can.

Evaluation

No published evaluation scores yet.