Social Draft: On-Device AI Social Reply Assistant, Making Every Response Just Right

Social Draft is an iOS on-device-first social reply assistant app, core to solving the awkward anxiety users feel when not knowing how to respond to messages. It positions itself as a "reply advisor" rather than a chatbot, with the final reply decision always in the user's hands. Through local large model inference, distilled SFT training, and user-controllable LoRA personalization, it provides natural and thoughtful reply suggestions. Meanwhile, its on-device-first design ensures privacy—data never leaves the device in local mode.

The product targets common awkward scenarios in real conversations: declining invitations when tired, politely refusing unwanted events, appropriate responses in emotional dialogues, concise professional replies for work/school, difficulty expressing oneself, etc. Unlike AI products on the market that "speak for users", Social Draft chooses to "help users speak", ensuring users retain the final decision on replies.

It uses a three-layer backend architecture: Mock (default safe mode, no model/API required), Cloud (calls APIs like OpenAI/Anthropic/Gemini), Local (on-device GGUF model, privacy-first). The Local mode is core—no network needed, data runs locally. It supports users training LoRA adapters locally to achieve style personalization, privacy protection, and lightweight adaptation. Currently supports Llama-3.2-1B/3B-Instruct-Q4_K_M and the reply_sft_lora_v1 adapter.

Provides multiple styles of smart suggestion cards (natural, direct, friendly, thoughtful, decision-oriented); Ghost Text inline completion feature (runs locally to speed up replies); context awareness (reads recent conversations to understand tone and topic); reply target selection (generates suggestions for specific messages).

Includes a two-stage Claude distillation process to generate synthetic social reply datasets; provides complete LoRA/SFT training notebooks to support users training personalized models with their own dialogue data; tools in the Experiments_Benchmarks directory can compare reply quality between local and cloud models.

The app uses SwiftUI to build a layered architecture (AppRootView, Features/Chat/Settings, etc.); local inference is implemented via llama.xcframework (loading GGUF models, mounting LoRA, sampling chain optimization); cloud supports OpenAI/Anthropic/Groq, etc.; Supabase integration is used for data persistence of demo chat services (thread management, real-time events, etc.).

Core functions run locally; users control reply rights; three modes are transparently distinguished; only necessary context is sent to the cloud. It avoids the ethical risk of AI replacing social interaction, positioning itself as an auxiliary tool to help users express themselves better.

Social Draft demonstrates the transformation of large model technology into a product that solves user pain points. Its advantages include on-device priority, LoRA personalization, and emphasis on privacy. For users, it alleviates social anxiety; for developers, it is a worthwhile on-device AI project to study, including complete application implementation and research workflow.