Chiblu Seller Outreach & AI Sales Agent — The Making of

The problem

The bootstrap problem was brutal.

Chiblu is India's curated marketplace for independent makers. Our buyers were on Instagram. Our sellers were on Instagram. I was one person.

The job description looked like this:

Find creators across 7 niches — clay, embroidery, jewelry, candles, art, miniatures, wooden goods
Scrape profiles to qualify them (followers, bio, originality)
Write 300+ personalized cold DMs a week
Reply to every "hi" within minutes
Handle objections — GST, commission, trust, "why should I switch?"
Track every lead through a funnel
Not burn out

Traditional outreach tools don't work in India. No one clicks links in cold DMs. No one trusts copy-pasted templates. And hiring an SDR at seed stage is how founders die broke.

The reframe

What if outreach didn't need to be human?

What if the seller outreach team didn't need to be human?

Not a drip campaign. Not a chatbot. A real system that:

Finds the right creators — not just followers, actual handmade makers
Writes DMs like I would — warm, founder-voiced, banned words: "artisan," "karigar," "handicraft"
Takes over the moment they reply — with full context, brand voice, objection handling
Hands back to me — only when a human judgment call is actually needed

What I built

Two modules. One system. One autonomous agent.

Module 01

The Outreach Dashboard

Batch Instagram DM automation. Reply-first, zero-link cold outreach.

Paste usernames → Apify scrapes → Claude writes DM → I hit send → Status tracked

12 playbook templates auto-recommended by niche + follower count + conversation stage
Niche auto-detection from bio keywords — clay vs. embroidery vs. jewelry gets different angles
Collab vs. seller pitch routing based on follower threshold (30K+ gets the collab angle)
Inline editing for notes, next action, creator name — no modals, no friction

The Seller's List — every creator, scored, tagged, and ready to pitch.

Module 02

Chiblu Assistant

Claude Sonnet 4.5 hooked directly into Instagram Graph API webhooks.

Creator replies → Meta webhook → Signature verify → Profile enrich → Claude responds → Reply in 1–4s

What makes it real, not a chatbot:

Message batching

If they send 3 messages in quick succession, the bot waits 5s, then replies once with full context — not 3 times.

Brand voice guardrails

Never says "artisan" or "karigar." Mirrors Hindi / Hinglish / English. Refuses to write code or recipes.

Anomaly detection

Rapid-fire, spam, 500+ char pastes get static replies — zero Claude tokens burned.

Human handoff

"Talk to a Human" auto-pauses the bot. I reply manually, then resume. Zero lost context.

Smart follow-ups

Cron job picks up stale leads daily, sends context-aware follow-ups — max 2 per lead, never spam.

Objection playbook

GST, commission, shipping, low-follower anxiety — Claude knows the answer, never makes one up.

The stack. Simple. No over-engineering.

Next.js 16

frontend + api

Supabase

postgres + rls

Claude 4.5

voice + logic

Meta Graph

dm + webhooks

Vercel

host + cron

Apify

profile scrape

One developer. One branch. Deploys on every push to main.

What it does

What it replaces. What it costs. What it produces.

What this replaces

Full-time SDR₹5–10L/yr
CS agent standby₹3L/yr
HubSpot Starter$20/mo
Outreach tool$100+/mo

What it costs me

Vercel Hobby₹0
Supabase free₹0
Claude API~$8/mo
Apify scrape~$2/500

What it produces

24/7 response time

1–4 seconds on Instagram DMs. While I sleep.

Auto-tagged leads

Every lead classified by language, city, product, intent.

Full conversation history

Searchable, exportable. Never lost in a DM scroll.

One dashboard view

I open it once a day. Not a full-time job.

1000s of creators.
Zero SDRs.