# I built an AI CRM where leads score themselves — 5 services on Zerops in 48h

> Source: <https://dev.to/sharmachaitanya945/i-built-an-ai-crm-where-leads-score-themselves-5-services-on-zerops-in-48h-ojd>
> Published: 2026-08-09 14:14:16+00:00

For the **#ZeropsChallenge** I built **LeadPilot** — an AI lead-automation CRM where every lead is scored, qualified, and given a drafted reply the instant it arrives.

Inbound leads pile up unqualified. LeadPilot fixes the triage: a lead hits the form → it's queued → a worker scores it 0–100 with Claude Haiku (intent, temperature, summary, personalized reply) → it lands in a live Kanban pipeline. Capture is instant; scoring is async.

| Service | Type | Role |
|---|---|---|
`app` |
static | React + Vite + Tailwind SPA |
`api` |
|

`worker`

`db`

`cache`

The API enqueues to Valkey and returns immediately; the worker (a separate service) does the slow AI work. Queue-decoupled by design.

Services talk over a private network via env references — no secrets in code:

```
run:
  envVariables:
    DB_HOST: ${db_hostname}
    DB_PASS: ${db_password}
    REDIS_HOST: ${cache_hostname}
    REDIS_PASSWORD: ${cache_password}
  start: python -m uvicorn app.main:app --host 0.0.0.0 --port 8080
```

Then zcli push api / worker / app. Migrations run once per version with zsc execOnce; /api/health gives zero-downtime rollouts.

AI usage (disclosed)

Scoring uses Claude Haiku with a deterministic fallback so it never breaks. I built it with Claude Code assistance — I designed the architecture and reviewed/tested everything.

Thanks @WeMakeDevs and @zeropsio! Try it: [https://app-2d82.prg1.zerops.app](https://app-2d82.prg1.zerops.app)