Akradan (AIX)
A full-stack AI collaboration platform that unifies access to 300+ LLMs in one workspace — with project-based context, knowledge-base RAG, document processing, cost tracking, and team sharing. Built for professionals and teams who need transparent, cost-tracked AI at scale.
Overview
Akradan (AIX) is a full-stack AI collaboration platform that unifies access to 300+ large language models through a single workspace. It adds project-based context and per-project knowledge bases (RAG), a document-ingestion pipeline, real-time team collaboration, an AI-driven business-tool suite, and Stripe subscription billing with per-model cost tracking. Built with Next.js and React Native on Supabase, with multi-tenant isolation enforced through PostgreSQL Row-Level Security.
Key Features
- Unified chat over 300+ AI models (via OpenRouter) with real-time token count + per-message cost, and a resilient fallback chain (Gemini Flash → Mistral → Phi 3) on model failure.
- Project-based workspaces with custom agent instructions and per-project knowledge bases — RAG via chunked embeddings, iDrive e2 S3 storage, and OCR fallback (Tesseract.js).
- Document ingestion accepting PDF, DOCX, XLSX, images, and text, with VirusTotal pre-scan, chunk indexing, and full-text retrieval injected into model context.
- Real-time collaboration: conversation sharing with invite tokens, pinned messages, 90-day retention (indefinite for pinned), and project ownership/access controls.
- AI-driven business-tool suite: contract management, PDF reports, pitch-deck builder, brand style guide, and social-media carousel creator.
- Stripe subscription billing (3 tiers) with a per-model usage ledger, credit reconciliation against OpenRouter balance, and configurable cost-warning thresholds.
Challenges & Solutions
Tracking real costs across 300+ models with wildly different per-token pricing required a dual-source reconciliation system: a local usage_charges ledger (written per response) cross-referenced against OpenRouter’s /v1/generation endpoint, with UI cost warnings and budget gates to prevent overruns.
Large RAG files (PDFs, DOCX, images) exceeded Next.js body limits, so uploads go directly to iDrive e2 via presigned S3 URLs, with VirusTotal scanning before indexing.
Long-running model calls (deep research, reasoning models) required switching API routes to the Node.js runtime with a 300-second execution cap and stream-based delivery to avoid timeouts.
Multi-tenant data isolation across chat history, knowledge bases, and shared projects is enforced entirely through PostgreSQL Row-Level Security policies — keeping all data-access logic at the database layer.
Results & Impact
?? Add real outcomes (users, MRR, retention) when available.