Welcome to Zedinga AI: Why we built an AI automation platform for SMBs

I run a small team. We build software for other small teams. So we know exactly what it feels like to evaluate a piece of "AI automation" and find out it's priced for a Fortune 500, locked to one industry, or so abstract you spend a weekend trying to figure out what it actually does.

That's the gap Zedinga AI is built for.

What this platform does today

Right now Zedinga ships service one of three: an AI customer support bot you can drop on any website in five minutes.

You give it your website URL, a few PDFs, or some Q&A pairs. It crawls, chunks, and embeds your content into a vector store. When a visitor asks something, it retrieves the most relevant chunks and writes a grounded answer using GPT-4o-mini — no hallucinations, no made-up facts, no answering questions you didn't equip it for.

Every piece is owner-controlled:

• Knowledge sources — web crawl, PDF upload, manual Q&A
• Persona — pick from five presets or write your own system prompt
• Languages — 12 supported with auto-detect
• Channels — web widget, WhatsApp Business, click-to-call buttons, live human handoff
• Plans — start free at $0/mo, scale to $199/mo without "contact sales"

If you're an SMB owner reading this and a single piece of the above sounds painful, the rest of the post might be useful.

What we deliberately did not build

A B2B enterprise sales motion. No demo bookings, no quote calls, no "starts at $5k/mo." Self-serve everything.

A walled garden of integrations that only works with one CRM. Webhooks, a public REST API, and Zapier compatibility — bring your own stack.

Closed-source telemetry. The repo is private but the data model is yours. Export your conversations, leads, and analytics any time as CSV.

A 6-month deployment. Most of our paying customers are productive on the same day they sign up.

The technical choices, in plain language

A few decisions that shape what Zedinga feels like:

1. Retrieval-Augmented Generation, not fine-tuning. Your data doesn't get baked into the model. Replace a knowledge source, the bot adapts immediately. Delete it, the bot forgets.
2. PostgreSQL + pgvector. No separate vector DB to manage, no ETL pipeline. Same database your account lives in.
3. Streaming token-by-token. Visitors see the answer forming in real time. Drops perceived latency by ~70% vs waiting for the full reply.
4. Shadow DOM widget. Embed code is one line. The widget can't accidentally break your site's CSS, and your site can't accidentally break the widget.

Where we're going

Two more services planned, both in the same self-serve, transparent-pricing spirit:

• AI Lead Generation — LinkedIn-first prospecting + CRM enrichment
• AI Business Process Automation — email triage, document processing, recurring reporting

I'll write a real post on each as we ship it. No vaporware countdowns.

Try it

The free plan covers 50 conversations a month, one bot, three knowledge sources. No credit card. If it works for your use case, upgrading is one button. If it doesn't, you'll have spent five minutes finding out — much better than a six-week pilot.

Thanks for reading. The blog will be a mix of product updates, technical deep dives, and honest "here's what we tried, here's what worked" posts. Subscribe via the RSS feed if that's still your thing.

— Sezgin