Collybrix GTM OS
Collybrix's LinkedIn prospecting was spread across LinkedIn tools, spreadsheets, Apollo and the Milano CRM. GTM OS brings it into one dashboard: it imports campaigns and replies, uses AI to suggest what each reply means, applies the sales plan's rules, puts whatever needs a person in a "Today" inbox and sends only interested leads on to the CRM. Internal tool, not publicly available.

The problem
At Collybrix most prospecting happens on LinkedIn, through automated campaigns run with LinkedIn tools. The data ended up scattered: invitations and replies in those tools, enriched contacts in Apollo exports, follow-up in spreadsheets and opportunities in the Milano CRM. Nobody knew for sure which reply was still unanswered, which campaign worked best or how many meetings came out of every hundred invitations.
What I built
- A dashboard with the LinkedIn funnel by cohort (invited, accepted, replied, interested, meetings), weekly activity, a per-campaign comparison and visible targets: pending invitations under 100, replies answered within one business day and AI accuracy.
- A "Today" inbox: what needs a person now (replies to classify, overdue actions, interested leads not yet in the CRM), with days counted in Spain's time zone.
- A lead view with the reply, the AI suggestion, the classification, the next action with its due date and the full history.
- AI reply classification in suggestion mode: Claude proposes the reply type, its confidence, the offer that fits and the date if one is mentioned. It is accepted with one click, and the human choice is stored next to the AI's to measure accuracy.
- Imports of backups and exports from the LinkedIn tools and Apollo CSVs, plus per-campaign webhooks with their own token.
- Sending to Milano: a deduplicated contact, a note per reply, a task with the deadline and an opportunity only on explicit interest. If the email is missing, it looks it up in Apollo by LinkedIn URL.
- Campaigns, services and projects: each campaign carries its LinkedIn account, segment, offer and variants, with its own metrics.
- A guided tour and a checklist that ticks itself from the data, so anyone on the team can use it.
Technical decisions
- Business rules in code: the reply type sets the stage, the next action, the deadline and whether an opportunity is created. A lead never moves back a stage and manual edits are never overwritten.
- Idempotent imports: re-importing the same backup, or a newer one, only adds what is new.
- The LinkedIn inbox never leaves the computer: the tool's backup is a SQLite database read in the browser with sql.js. Only the people contacted by the chosen campaigns, and their replies, are sent.
- GDPR by design: a daily job anonymizes prospects with no activity for 30 days and keeps the metrics. Opt-outs are kept only as a keyed hash, so no import can re-create them, and webhooks store a minimal copy with no phone numbers or message history.
- Access only through the company's Google Workspace: self-hosted Better Auth, revocable sessions stored in the database, and every page and route validates the session, not just the proxy.
- AI through Vercel AI Gateway with zero data retention.
Status
In internal use at Collybrix. It started in September 2026 as a lead log with Apollo imports, and in November it gained the LinkedIn tools integration, AI classification, metrics and sending to Milano. The screenshots use fictitious data.


