GhostReply
An AI assistant that finds the conversations worth joining and replies in your own voice
THE CHALLENGE
Creators and founders struggle to stay consistently active on social media. Finding relevant conversations takes time, writing thoughtful replies takes effort, and generic AI tools produce answers with no personal voice or context. GhostReply combines personalized discovery with reply generation based on how the user actually writes, and keeps the user in control of what gets published.
TECH STACK
Engineering Approach
One connection model, many platforms
Social accounts connect through a shared internal model, with a separate adapter for each platform's API. X comes first, and LinkedIn, WhatsApp Business, Instagram and Facebook follow without touching the AI layer.
A voice profile instead of a generic prompt
Each user saves their tone, personality and writing preferences. Every reply is generated from that profile plus the context of the post, so the output sounds like the user rather than a template.
Discovery built around topics
Users pick the topics they follow. The system retrieves matching public posts, filters duplicates, ranks them by relevance, and serves them in a paginated feed of 10 posts per page.
Review before anything goes out
Generated replies are drafts. The user edits, regenerates or approves them, and only approved replies are published through the platform's official API.
Key Technical Decisions
The choices that shaped the architecture and how easy it is to maintain.
Official platform APIs instead of a unified provider
Each platform integrates through its own official API and OAuth flow, not through a paid third-party social media integration service.
This avoids a per-connection cost and a dependency on a vendor's coverage and pricing, and it gives full control over permissions, token handling and rate limits. The cost is building each integration, which the shared connection model is designed to keep small.
OpenRouter for reply generation
Reply generation goes through OpenRouter, behind an AI layer that does not depend on any social platform.
One API keeps the choice of model open, so quality and cost can be tuned without changing product code. Keeping the AI layer separate from the integrations means adding a platform never changes how replies are generated.
Human approval before publishing
Nothing is published automatically. Every reply is shown to the user for review first.
Replies go out under the user's own name, so a mistake is public and personal. Requiring approval keeps the user in control of their voice and their reputation.
Cached discovery with controlled refresh
Discovery results are cached and refreshed on a controlled interval, with rate-limit handling and token refresh in the integration layer.
Platform APIs are rate-limited and cost-sensitive. Caching keeps the feed fast and API usage predictable, and background jobs keep results fresh without calling the API on every page load.
