Built for
A production-grade AI chatbot platform. Multimodal, multi-channel, zero-downtime model switching. Ship in days, not quarters.
Live demo
Voice note in.
Verified receipt out.
Real conversations, not mock screens. The bot transcribes the voice note with Whisper, parses the QRIS with PaddleOCR + vision, and writes back in a voice that sounds like a human on the other end of the line.
- Voice note transcription (Faster-Whisper v3)
- QRIS & receipt OCR (PaddleOCR + Qwen vision)
- Image understanding + intent routing
ONEXVIBES AI
online · typically replies instantly
“Transfer 100rb ke Budi, QRIS di foto kedua”
Oke kak, QRIS Rp 100.000 untuk Budi Santoso terdeteksi. Konfirmasi transfer ya?
Capabilities
Everything a modern chatbot needs,
already wired together.
22 layers, 195+ components. Opinionated where it matters, swappable everywhere else.
Multi-channel out of the box
WhatsApp, Telegram, Instagram, Facebook, and your web widget — one message bus, one memory, one persona.
Voice, vision & documents
Transcribe voice notes, read receipts, decode QRIS, understand stickers. Users talk how they want.
Zero-downtime model swap
Ship a new model without dropping a session. Gradual traffic shift, automatic rollback, live metrics.
Model-agnostic brain
Swap Qwen for Claude, Llama, or any OpenAI-compatible endpoint. Persona, tools, and flows stay put.
Enterprise-grade security
Cloudflare WAF, field-level encryption, KYC/AML ready, full audit trail. Built for regulated industries.
Observability from day one
Langfuse traces, Grafana dashboards, cost per tenant, token per conversation. You see what the bot sees.
Human-like by design
Persona consistency, typing delays, message chunking, sentiment-aware replies. Users forget it's a bot.
Function calling that ships
Connect to payments, CRM, balances, promotions. The model decides which tool; you write the TypeScript.
OCR tuned for Indonesia
Bank slips, QRIS, e-wallet receipts — parsed with PaddleOCR + vision, not guessed.
How it works
Three steps to a chatbot you'd actually ship.
Plug your channels
Connect WhatsApp, Telegram, Instagram, or drop the web widget. A unified inbox, one message schema.
Define persona & tools
Write your system prompt. Declare function-calling tools — balance checks, payments, CRM lookups. Typed.
Pick a brain, ship it
Qwen 3.6 on your GPU, Claude on cloud — toggle from the panel. Zero downtime, automatic fallback chain.
Full production stack, end to end
Pre-wired, priority-tiered
Text, voice, vision, sticker, QR…
Warm standby & auto failover
The stack
Boring tech where it should be boring.
Sharp where it matters.
AI Stack
- Qwen 3.6-35B-A3B FP8
- SGLang 0.5.10
- Faster-Whisper v3
- BGE-M3 + Reranker
- NeMo Guardrails
- LiteLLM
Backend
- Node.js + Fastify
- TypeScript
- XState
- BullMQ
- PostgreSQL 16
- Redis Cluster
- Qdrant
- Kafka
- MinIO
Channels & CX
- Evolution API (WA)
- Telegram Bot API
- Instagram DM
- Facebook Messenger
- Chatwoot
- n8n
Multimodal
- Sharp + ImageMagick
- PaddleOCR
- ZXing
- FFmpeg + RNNoise
- Silero VAD
- ClamAV
Security & Ops
- Cloudflare WAF
- HashiCorp Vault
- Keycloak
- Langfuse
- Prometheus
- Grafana
- Docker
Compliance
- KYC (Sumsub)
- AML monitoring
- GDPR toolkit
- Immutable audit log
- Self-exclusion
- Usage limits
FAQ
The questions every team asks.
Do I need my own GPU?
No — the stack runs on cloud APIs (Claude, GPT) out of the box. If you want on-prem control and lower per-token cost at scale, an RTX PRO 6000 or H100 slot is plenty.
Which channels are supported today?
WhatsApp (Evolution API + WABA), Telegram, Instagram & Facebook DM (Meta Graph), embeddable web widget, email, and SMS OTP. TikTok DM is behind a feature flag.
Can I swap the model without losing context?
Yes. Conversation history, memory tiers, tool definitions, persona, and XState flows are model-agnostic. The 'brain' is the only external thing — flip the switch, traffic shifts gradually.
Is this compliant for regulated industries?
The Compliance layer ships KYC (Sumsub), AML monitoring, jurisdiction gating, self-exclusion, and immutable audit logs. You still need your own license — we don't sell one.
How fast can I ship?
Pilot on a single channel: 1–2 weeks. Full production with multimodal + admin panel: 4–8 weeks depending on your business integrations.
What about my existing cPanel site?
Everything non-AI runs in Docker on Server 1 alongside cPanel. Your website keeps its slot. The GPU work lives on Server 2, reachable only on a private network.
Stop gluing libraries.
Start shipping conversations.
A pragmatic, opinionated AI chatbot stack — ready for your first pilot this month.