OPERAND STUDIO
RUNNING · INTERNAL · COMMUNITY LEAD RADAR

Radar — the right conversations, surfaced, never automated away.

A social-listening system on n8n: it watches developer and business communities, LLM-qualifies each new post for genuine intent, and pings the owner with a short brief and a value-first angle. Detection is automated; every reply is written by a human. This is the system we run for ourselves.

SITUATION

The best leads are people asking, in public, right now.

Across communities like Hacker News and the n8n forum, people describe exactly the problem you solve — asking for help, evaluating tools, hiring. Answering the right one, early and usefully, is how a small studio earns trust. But the good posts are buried in noise and gone by the time you scroll past them.
BOTTLENECK

Manual monitoring doesn't scale — and auto-reply gets you banned.

The two obvious approaches both fail:
SYSTEM

A scraper base, an LLM qualifier, two delivery tiers — one human at the end.

A scraper polls each source through its official API, dedupes against a seen-table, and advances a durable per-source cursor so it only ever fetches what's new. A cheap keyword pre-filter cuts the obvious noise before any model runs. What survives goes through one guarded LLM pass that scores relevance, intent, fit, and actionability into a composite, with a short reason and a suggested value-first angle. High-scoring leads ping the owner live on Telegram; everything else collects into a daily email digest. A human reads the brief and writes every reply.
The Radar system: detection automated, response human. A scraper polls official community APIs — Hacker News via Algolia, the n8n forum, with Reddit deferred — deduping and advancing a per-source cursor. Surviving posts pass a keyword pre-filter, then dispatch fail-quiet to an enrichment step: one JSON-only, spend-guarded LLM pass scoring is-relevant, lead type, fit, intent and actionability into a composite, with a reason and a value-first angle. Leads at or above the threshold ping the owner live on Telegram, with Attended, Deferred and Mute controls; everything else routes to a daily email digest that doubles as a searchable archive. A human writes every reply — the tool never posts anywhere.
THE PIPELINE, IN ORDER
  • Scrape: each source is polled through its own structured API (Hacker News via Algolia, the n8n forum via Discourse). A durable per-source cursor means each poll fetches only newer items; a seen-table dedupes as a safety net.
  • Pre-filter: a deterministic keyword pass (ask-signals, fit terms, negatives) drops obvious noise before any model spend.
  • Enrich: one LLM pass, JSON-only, returns is-relevant, a lead type, and fit / intent / actionability sub-scores. Showcases, off-topic, and pure discussion are hard-gated out.
  • Score: sub-scores fold into a composite, adjusted for budget and urgency signals, then bucketed into tiers.
  • Deliver: top-tier leads ping the owner on Telegram with a brief and a value-first angle; the rest land in a daily email digest — a searchable archive, not a flood.
  • Respond: the owner reads the brief and writes the reply. One-tap Attended / Deferred / Mute feedback records the outcome; the angle proposes how to genuinely help, never a scripted product pitch.
Detection automated, response human

The radar finds and qualifies. It never posts, comments, or messages a community — every reply is written by a person. Automating the reply is a spam-and-ban trap, out of scope by design.

Official APIs, not scraping

Sources are polled through their own structured APIs — Hacker News via Algolia, the n8n community forum via Discourse. Durable per-source cursors fetch only what is new; a seen-table dedupes.

Fail-quiet, spend-capped

If enrichment is down or absent, posts route to the digest — never a flood of raw pings. The LLM pass is JSON-only and guarded by a hard daily spend cap with an alert.

OUTCOME

Measured, not promised.

Every qualified post is stored with its full scoring — sub-scores, composite, the reason, and the one-tap feedback the owner leaves. That record is the measurement instrument; the outcome section publishes from it, not from claims.
RUNNING · INTERNAL

The radar runs on our own lead-gen. As real feedback accrues, the numbers land here: posts surfaced by source, precision by tier (pinged then actioned), and the share caught that manual scanning would have missed — pulled straight from the stored feedback. Until the window closes this section stays deliberately empty. No invented figures, ever.

PACKAGES

A base scraper other packages stack onto.

The scraper is the base brick — source monitor, dedupe, cursor, deployed per target as-is. Enrichment, the daily digest, and the eventual CRM push are add-on packages on fail-open seams: each is independently useful and reusable on other bricks.
Radar — scraper (BASE)Source monitor over official community APIs — dedupe, durable cursors, keyword pre-filter. Runs alone by design.BASE
Radar — LLM enrichment (ADD-ON)One JSON-only, spend-guarded LLM pass that qualifies and scores each post, with a value-first angle.ADD-ON
Radar — daily digest (ADD-ON)Composes the review-band leads into a daily email — a searchable archive that doubles as the low-urgency delivery tier.ADD-ON
Talk about finding your leads →View the components on GitHub →