A sales performance AI for a B2B conferences operator

A global B2B conferences operator runs outbound across 70+ sales reps distributed across three continents. Activity sits in five disconnected systems — Twilio for calls, Zoho for CRM and deals, Apollo for cold email, Insightful for activity tracking, Outlook for direct mail — plus an Excel revenue tracker that lives outside any CRM. At that scale, manual review of even a fraction of weekly calls is impossible. Individual rep performance is invisible, and coaching defaults to gut feel.

• 01Built a custom manager-facing web app on Next.js, React, and Postgres, with Microsoft 365 SSO as the auth gate.
• 02Wired the Twilio call-recording pipeline — server-side audio proxy so managers stream any call straight from the rep page, no credentials exposed.
• 03Stood up an automated transcription pipeline (ElevenLabs Scribe v1, diarised) — per-call from the UI or as a nightly batch with concurrency caps.
• 04Wired ingestion from Zoho CRM (users, calls, deals), Apollo (sends, deliverability, opens, clicks, bounces), Insightful (PC-tracked activity time), and Outlook (direct mail). One reconciled view per rep, colour-coded by source.
• 05Built a Gemini Flash call-scoring engine against a configurable nine-criteria rubric — call-type-aware so gatekeeper hits and voicemails don't pollute rep averages. Rubric is admin-editable and versioned.
• 06Built a second AI layer that reads the last 30 days of transcripts + scores per rep and writes a structured weekly coaching review — strengths, recurring themes, one coaching priority per week. Explicitly non-judgmental: doesn't punish high-call/high-idle reps, doesn't blame missing channels on the rep.
• 07Built a closed-loop coaching surface: flagged calls land on a 'Coach Today' worklist, manager opens the call, marks it reviewed, the call drops off the list and a reviewed badge appears on the rep's coaching feed.
• 08Built a deal-attribution stack — drag-drop Excel parser handling five row patterns (single, multi-attendee, sponsor-multi-event, cancelled, refunded) with idempotent re-uploads, plus an admin route to reconcile informal first-name references to specific reps.
• 09Built admin/employee role separation: reps see their own strengths, themes, and weekly coaching priority — never the counted-failure language, individual scores, or peer rankings.

The client now monitors a 70+ rep distributed sales floor at a level of detail that was impossible before.

Every recorded call is transcribable on demand or via nightly batch. Every scoreable call gets a structured grade with tags. Every rep gets a weekly AI-synthesised coaching priority drawn from patterns across their own conversations. Managers see ranked top and bottom performers, a 'Needs coaching' filter, and a closable Coach Today worklist. The build is active and in production; cross-channel data continues to thicken as live collectors come online.