Okidoki.chat: A Year of Building an AI Conversion Platform
From a 3-week MVP to text, voice, video and WhatsApp agents booking real meetings for 213+ businesses
Where It Started
Okidoki.chat launched as a 3-week MVP: a fast RAG chat widget with sub-200ms responses, voice mode, and video calls. It proved the concept — brand-specific answers could feel instant instead of "AI slow" — and it landed the first paying customers.
But here's the honest part: speed is not a moat, and it's not a product. A widget that answers questions well is a demo. What businesses actually pay for is outcomes — qualified leads, booked meetings, less time wasted on tire-kickers.
The last year of work has been about closing that gap. This article is about what that took.
What It Is Today
Okidoki.chat today is an AI conversion platform: it engages website visitors by text, voice, real-time video avatar, or WhatsApp, answers from the business's own knowledge, qualifies the lead, and books the meeting — automatically, 24/7.
What shipped between the MVP and now:
- Four channels, one brain — the same AI agent and knowledge base serve text chat, voice conversations (30+ voices), face-to-face video with lifelike avatars (20+), and WhatsApp
- Smart scheduling — the agent qualifies visitors and books directly into real calendars, with video meetings (HD, screen sharing, automatic transcription and AI summaries) built in
- AI Page Copilot — the agent can interact with the host page itself: fill forms, navigate, guide the visitor
- Human handoff — when a visitor needs a person, the team is notified instantly (push, web, email) and can take over from a dashboard or the mobile Operator app, with full conversation context
- Lead intelligence — automatic visitor profiles, engagement history, and ICP profiling
- Multi-tenant workspaces — teams, multiple widgets per account, white-label branding, AI-generated landing pages and immersive 3D sites
- An
ask()API — so the same brain can power integrations beyond the widget
Live: okidoki.chat↗ — and running in production on this website right now.
The Hard Problems
The MVP's hard problem was RAG latency. The platform's hard problems were different — and each one took longer than the entire original build:
1. One brain, four channels. Text tolerates a 500ms pause; a voice conversation doesn't. Making the same knowledge base and agent logic feel native in chat, in a live voice session (Gemini Live), on a video avatar with lipsync, and inside WhatsApp's messaging model meant rethinking the pipeline per channel while keeping a single source of truth.
2. Real-time avatars that don't break trust. A video agent that lags or glitches is worse than no video agent. Getting avatar rendering, speech, and interruption handling to feel like a person required aggressive latency budgeting across every hop.
3. Handoff is a product, not a feature. The moment AI hands a conversation to a human is where deals are won or lost. That meant notifications that actually reach people (push, web, email), a mobile app for responding anywhere, and handing over context, not just a transcript.
4. Knowledge that stays current. Businesses change their sites and documents constantly. Ingestion had to become continuous — website crawling, document processing, and re-indexing on a schedule — so answers never drift out of date.
5. Multi-tenancy and reliability. 213+ businesses means real SLAs: 99.9% uptime, GDPR compliance, per-tenant isolation, usage metering across plans. Unsexy, invisible, and where most of the engineering year actually went.
Architecture
The original performance philosophy survived and scaled: do the heavy work before the question arrives.
- Build-time RAG processing — content is indexed when it changes, not on every query
- Redis — processed embeddings and hot query patterns close to the edge
- Groq for low-latency inference, heavier models where accuracy matters
- Vercel Edge Functions — no cold starts, global deployment
- Gemini Live for real-time voice, Daily.co for WebRTC video, AssemblyAI for transcription
What a Year in Production Taught Me
Speed is a starting advantage, not a value proposition. The 3-week build was great for validation — and irrelevant to why customers stay. They stay because meetings show up on their calendar. I stopped selling "built fast" and started measuring booked meetings.
The MVP is 10% of the product. Everything that makes Okidoki worth $149–$999/month — handoff, WhatsApp, lead profiles, reliability, the mobile app — came after launch, driven by watching real businesses use it and fail at specific moments.
Retention lives in the boring parts. Nobody churns over a missing feature announcement; they churn over a missed notification or a stale answer. Continuous ingestion and bulletproof handoff did more for retention than any flashy capability.
Dogfooding compounds. This website runs Okidoki, and my own consulting leads get qualified and scheduled by it. Every rough edge I feel as a user becomes next week's fix.
Try It Yourself
This website uses Okidoki.chat in production. Open the widget (bottom right) to:
- Ask questions about my experience — answers come from my own content
- Try the voice mode
- Let it book a meeting with me, end to end
Building an AI product past the demo stage? That journey — from prototype to platform — is exactly what I help teams with. Let's talk.