All work

    AI call center CRM with voice agents

    AI voice agents dial your leads, every call is transcribed and scored hot/warm/cold, and an AI assistant flags stalled deals before you lose them.

    Built for: CloserIntelligence · Updated 2026-09-01

    Who it's for

    Sales teams and call centers running outbound at volume.

    The problem

    The client was running a call centre on a CRM that recorded activity but never acted on it. Reps logged calls, deals stalled, and nobody found out until the pipeline review. They needed the system to detect the stall and act on it, not report it a week late.

    What it does

    • 01AI voice agents make outbound calls unattended.
    • 02Every call is recorded, transcribed, and classified by sentiment and stage.
    • 03A conversations inbox scored hot/warm from the transcript, not a manual tag.
    • 04Campaigns with dial volume and connect rate per campaign.
    • 05Knowledge bases the voice agents read from.
    • 06Maclovi, an AI operations assistant that reads the call log and raises tickets on stalled deals.
    • 07A prompt optimizer to tune agents against real call outcomes.
    • 08Multi-tenant, with a superadmin portal.

    Building something in this shape? Book a 30-min scope call →

    The product

    These are the authenticated screens, not the marketing site. Lead names, home addresses, phone numbers and the client's own contact details are blurred — it's a live workspace running real outbound, and I'm not publishing third parties' details to make a nicer portfolio shot. Everything else is untouched.

    Dashboard — call outcomes across the workspace: 8,157 no-answer, 2,483 busy, 929 hung-up, 240 voicemail.
    Conversations — inbound threads scored hot/warm off the transcript, not a manual tag.
    Call logs — every dial with duration, outcome, recording, and the extracted transcript.
    Leads — 41,494 records with pipeline stage, source, and per-row call history.
    Campaigns — dial volume, connect rate, and agent assignment per campaign.
    Knowledge bases — the training corpora each voice agent reads from (compliance, auction, probate).

    Walkthroughs

    Developer POV — architecture walkthroughThe dependency graph, route map, and Supabase surface behind the product. This is the build, not the pitch.
    Product tour — inside the running workspaceEight screens from the live client workspace: dashboard, conversations, call logs, leads, agent directory, prompt optimizer, escalation alerts, and team management. Lead names, addresses and phone numbers are blurred — this is real call data, not a staged demo tenant.

    How it's built

    ReactTypeScriptSupabaseVAPIGeminiGoHighLevelReact QueryRadix UITailwind

    Role: Sole engineer — architecture, frontend, Supabase schema and policies, VAPI integration, GoHighLevel sync.

    What's in the codebase

    472
    Modules in graph
    madge dependency graph of the source repo
    138
    Dependency edges
    internal import relationships
    25
    Routes
    app-level pages incl. superadmin + invite flows
    60
    Frontend deps
    production package.json dependencies

    The build

    • 01Built the platform as a multi-tenant React + TypeScript app on Supabase, with 25 routes covering agents, campaigns, call logs, conversations, leads, knowledge bases, phone numbers, call quality, and a superadmin portal.
    • 02Wired VAPI voice agents into the call pipeline so outbound calls run unattended, and post-call transcripts land in Supabase with sentiment and classification attached.
    • 03Built Maclovi, the operations assistant layer — it reads the call log and raises tickets against stalled deals rather than waiting for a human to notice.
    • 04Layered a prompt optimizer route so the client can tune agent prompts against real call outcomes without redeploying.
    • 05Two-portal auth with internal and external invite flows, so the client's own team and their sub-accounts land in different surfaces from the same login.

    Engineering decisions

    Transcript classification stored, not computed on read

    Sentiment and stage classification are written once at ingest and stored as columns. Recomputing them on every dashboard render would have made the call log unusable past a few thousand rows, and would have burned model spend on data nobody was looking at.

    Badges as leaf components

    The most-imported modules in the graph are ConversationStageBadge, ClassificationBadge, and SentimentBadge. Keeping classification display in tiny leaf components meant changing the taxonomy was a one-file edit rather than a sweep across every table and detail view.

    React Query as the only server-state layer

    No Redux, no context soup. Every Supabase read goes through React Query with a keyed cache, so the war-room views can poll aggressively without stacking duplicate requests.

    Want your own AI call center CRM?

    Bring the problem, not a spec. 30-minute call — I'll tell you what it actually takes to build, what I'd cut, and what it costs. No decks, no discovery phase invoice.

    Related builds