How SignalPipe scores buying intent

1. The problem keyword alerts can’t solve

Suppose you sell a tool that automates Reddit lead discovery. A keyword alert for “Reddit leads” will fire on every post that mentions Reddit and leads, regardless of whether the author wants to buy something or sell something. The signal-to-noise ratio collapses fast.

A lexical match doesn’t know the difference between “I’m looking for a tool to find Reddit leads” (a buyer) and “5 ways our tool finds Reddit leads” (a competitor’s blog post). A semantic match does. That’s the entire reason SignalPipe exists.

2. Anchor sentences as semantic targets

Each product configures 5–10 anchor sentences: short, natural-language examples of the buying intent it wants to detect. They’re written in the buyer’s voice, not the seller’s. Examples:

• “I need a tool to monitor Reddit for sales leads”
• “Looking for an AI sales agent that can find prospects automatically”
• “My outbound process is too manual and I want to automate lead discovery”

Mantidae embeds each anchor with OpenAI’s text-embedding-3-small and caches the vectors at product-load time. Incoming RSS / Reddit / HN posts are embedded the same way. The cosine similarity to the closest anchor becomes the embedding component of the signal score.

Why anchors instead of a single product description? Because real buyers don’t describe your features — they describe their pain. Several short anchors covering different framings of the same intent outperform a single long product blurb on every dataset we’ve tested.

3. Multi-factor geometric mean

Embedding similarity is necessary but not sufficient. A post can match an anchor semantically but still be irrelevant — a stale repost, a comment from someone with no audience, a sarcastic reply. Mantidae combines five components into one weighted geometric mean:

• Embedding match — cosine similarity to the closest anchor
• Keyword density — buy-signal keywords found in the post
• Freshness — exponential decay on age, faster on social platforms
• Engagement — comment / vote count where available
• Author reputation — follower / karma signal where available

We use a geometric mean instead of a weighted sum because it punishes weak components more aggressively — a post with a great embedding match but zero engagement and a stale timestamp shouldn’t score as high as a fresher, more-engaged post. The geometric mean enforces that all signals matter.

The output is a 0–100 content score — the truth signal before any post-processing.

4. The competitor-floor heuristic and its honesty cost

If a post mentions one of your configured competitor names, Mantidae enforces a minimum score of 75. Why: someone publicly evaluating your competitor is one of the highest-intent signals there is, and you don’t want a weak embedding match to bury it.

The cost: this rule fires for any mention of the competitor — including the competitor’s own marketers talking about their own product. To stay honest about it, Mantidae preserves the pre-floor content_score alongside the post-floor signal_score. The dashboard surfaces both numbers and visually flags missions where the gap between them is ≥ 30 points.

Crucially, the role assignment for the drafting swarm uses content_score, not signal_score. So a misclassified competitor mention surfaces in the queue (good — operator should still see it) but the draft stays value-first instead of getting upgraded to a closing pitch (good — we don’t want to send hard CTAs to the competitor’s own marketing team).

5. Reinforcement-learning feedback loop

Each listening station (an RSS feed, a subreddit, an HN search) carries its own rl_weight (default 1.0, clamped 0.5–2.0). When a post comes in, Mantidae multiplies the raw signal score by the source station’s weight before the 50-point threshold. So the system can demote a single noisy feed without penalising the rest of the product’s sources.

The weight updates on every operator decision. Approvals are flat; rejections are reason-aware — the size of the penalty matches how unambiguous the failure mode is.

• Approve a mission → +0.05 (this station produces leads worth pursuing)
• Reject — spam → −0.04 (bot or promoted post — strongest signal that the feed is broken)
• Reject — not_relevant → −0.03 (wrong audience or topic for this product)
• Reject — no_reason / too_vague → −0.02 (default; weak signal)
• Reject — sarcasm / wrong_product → −0.01 (real signal, just not buyable — gentlest)
• Reject — already_customer → 0.00 (positive outcome misclassified as a lead — no penalty)

Approvals are larger than the typical rejection penalty for the same reason as before: false negatives (missing a real prospect) cost more than false positives (one extra mission to skim). Per-station weights mean RL behaves like a feed quality score that actually distinguishes which sources are working — not a single global multiplier that drags every station for the product down together.

6. The 9-prompt role-aware swarm

Drafting is handled by a 3-judge swarm: a Skeptic, an Analyst, and an Optimist, each running a different system prompt. They produce three drafts in parallel; we fuse them using the geometric mean of their self-scores and pick the highest-fused candidate.

The role-aware part: each lead’s content_score determines a role — closer (> 80), advisor (61–80), or educator (40–60) — and the role swaps in a different system prompt for each of the three judges. That’s a 3 × 3 = 9-prompt matrix. The voice differences:

• Closer — propose a specific concrete next step (demo, trial, link)
• Advisor — consultative, acknowledge the problem, introduce the product as a natural fit
• Educator — answer the question first, pitch only if it fits naturally

Char budgets are passed in too: 280 for Twitter replies, 500 for Reddit DMs, 300 for manual outreach. The swarm targets the budget at draft time — operators don’t inherit a draft that needs to be cut in half.

7. Failure modes we know about

Honest list. Most are mitigated; one or two are open.

• Sarcasm — embedding similarity can’t detect inverted intent. Mitigated by a sarcasm-detection unit-test pass on the sidecar before drafting.
• Stale reposts — same URL crossposted to 5 subreddits. Mitigated by the interactions ledger (unique on URL + product).
• Competitor-marketer false positives — discussed in §4. Surfaces in queue but doesn’t produce a hard pitch.
• Bot accounts and karma farms — open. Author-reputation component helps but doesn’t fully solve. Operator review is the backstop.
• Cross-product anchor leakage — a generic anchor like “I need a sales tool” will match too broadly. Best practice: write anchors as specific as the product warrants.