AI Business Core
White-label RAG platform for document Q&A. Businesses upload docs, embed AI assistant anywhere. My demo: chat with my CV on this site and Telegram.
The Problem
Businesses wanted AI assistants trained on their own documents (policies, FAQs, product info) without building ML infrastructure. Existing solutions were expensive enterprise tools or required AI expertise. I needed a multi-tenant platform where each business gets isolated data, embeddable widget, and zero hallucination answers.
Constraints
Strict tenant isolation - businesses must never see each other's data
No hallucination - AI must only answer from uploaded documents
Must work everywhere - web widget, Telegram, WhatsApp
One-click embed - customers paste a script tag, done
Technical Decisions
Frontend
- Built embeddable widget that works with one script tag
- Implemented streaming responses for real-time feel
- Designed business dashboard for document upload and conversation history
- Added customizable branding (colors, logo, welcome message)
Backend
- Built RAG pipeline: PDF/markdown parsing, chunking, OpenAI embeddings
- Used PostgreSQL + pgvector for semantic search across documents
- Designed multi-tenant schema with row-level security
- Created Fastify API with WebSocket support for streaming
Infrastructure
- Deployed on Railway with auto-scaling
- Set up S3 for document storage with tenant-scoped IAM
- Implemented rate limiting per tenant to control costs
- Added monitoring for embedding costs and query volumes
Trade-offs Made
- !
Chose pgvector over Pinecone/Weaviate to keep infrastructure simple
- !
Accepted per-tenant latency hit for stronger isolation guarantees
- !
Built custom widget instead of using Intercom/Crisp for full control
Outcome
Live platform with my 'Ibrahim' demo: this website's chat widget and Telegram bot (@ibra_cv_ai_bot) answer questions about my experience, projects, and skills - all from my uploaded CV. Businesses can create their own assistants with their own documents.
Interested in discussing this project or similar challenges?
Get in touch