OneAI is a high-performance, developer-first AI workspace that consolidates the world's leading Large Language Models into a single, unified interface --- without SaaS restrictions, rate caps, or platform lock-in.
It is designed as a personal power tool for engineers who want:
- Direct API-level control
- Intelligent multi-model routing
- Real-time web-grounded responses
- Strict cost governance
- Zero commercial abstraction layers
This project demonstrates system design across AI orchestration, RAG pipelines, secure key handling, and vector search infrastructure.
Supports:
- Gemini (2.5 Flash, 3 Pro)
- GPT‑4o
- Claude 3.5 (Sonnet / Haiku)
- DeepSeek (V3, R1)
- Grok (2, Beta)
Slot-Based Architecture - Map OpenRouter master keys or official provider SDK keys - Dynamically assign models per slot - Swap providers without backend reconfiguration
The AIFactory layer:
- Detects OpenRouter vs Official API keys
- Normalizes model identifiers
- Dynamically formats headers
- Prevents routing mismatch failures
Result: consistent model invocation regardless of frontend request format.
OneAI goes beyond standard chat interfaces by supporting Artifacts—dedicated, stateful workspaces for code, documents, and structured data.
- Stateful Execution: The backend continuously tracks the
current_artifactstate within the database for each specific chat session. - Iterative Refinement: Models can read the existing artifact state and iteratively update, patch, or rewrite code blocks without losing context.
- Cross-Model Handoff: Because artifact state is persisted at the database level per chat, you can generate a codebase with DeepSeek, seamlessly switch to Claude 3.5 Sonnet, and ask it to refactor the exact same artifact.
When temporal or real-time intent is detected, OneAI activates a custom Retrieval-Augmented Generation workflow:
- Search -- DuckDuckGo HTML query (non-sponsored results)
- Scrape -- Extract clean article content via Cheerio
- Chunk -- RecursiveCharacterTextSplitter (LangChain)
- Embed -- Vector embeddings generated per chunk
- Store -- Supabase pgvector
- Retrieve -- Cosine similarity top‑k match
- Inject -- Context + source URLs prepended to prompt
This ensures responses are grounded in current, external information instead of static training data.
- Real-time token tracking
- Model-specific daily budget ceilings
- Loop and runaway prevention
- No artificial SaaS paywalls --- just enforced safety thresholds
Demonstrates practical API governance design.
- Flutter (Dart)
- Riverpod
- Shared Preferences
- Material 3 (Adaptive theming)
- Node.js
- TypeScript
- Express
- Cheerio
- LangChain
- Supabase (PostgreSQL + pgvector)
This project highlights:
- Stateless API design with secure header key injection
- Vector similarity search with pgvector
- Custom RAG pipeline implementation
- Dynamic multi-provider orchestration
- Budget-aware AI execution model
- Clean separation between client key storage and backend execution
- API keys stored locally on-device
- Backend never persists credentials
- Stateless execution model
- No multi-tenant storage layer
- No remote key logging
Designed intentionally as a private AI control workspace.
OneAI is intentionally not a public SaaS product.
It is built as:
- A personal AI engineering lab
- A project to learn system design
- A daily-driver AI orchestration workspace
The absence of authentication layers and RLS is deliberate to maintain simplicity and control for single-user execution.
This repository demonstrates practical application of:
- AI orchestration patterns
- Retrieval-Augmented Generation systems
- API cost governance
- Secure key architecture
- Cross-platform frontend engineering
- Vector database integration
It is not just an AI wrapper --- it is an infrastructure-level AI control system.
Fun personal project. Not intended for commercial redistribution.