Crivox
The Problem
Posting consistently is only half of social media. Commenting thoughtfully on other people's posts is what builds visibility and relationships, but writing good comments at volume is slow, so most people either skip it or repeat the same short replies.
Why Existing Approaches Fall Short
- •Generic AI replies sound templated and often miss the point of the original post.
- •Every platform has its own tone and etiquette, and one style does not fit all of them.
- •Managing comments across many posts by hand means nothing gets tracked or scheduled.
What Was Built
- ✓Comment generation from pasted text, a shared link, or an uploaded screenshot.
- ✓Up to five variations for each post, across eight tone styles.
- ✓Bulk generation for several posts at once, with CSV export.
- ✓A queue for scheduling comments, tracking them and marking them as posted.
- ✓A template library with built in presets plus custom templates.
- ✓An analytics view covering activity trends, tone distribution and a breakdown by platform.
- ✓Saved history with search, filters and favourites so good comments can be reused.
- ✓Shareable result links that open without requiring a login.
- ✓A voice profile that stores writing samples so output matches the user's own style.
How It Works
Add the post
Paste the text, share a link, or upload a screenshot of the post.
Choose a tone
Pick one of eight tones and, if needed, a language for the reply.
Pick a variation
Review up to five options and keep the one that fits the conversation.
Track it
Add the comment to the queue and mark it as posted when it goes live.
Key Architectural Decisions
Read screenshots instead of asking for a copy and paste
A lot of engagement happens on a phone, where copying a post is awkward. Accepting an image directly removes the most common reason people abandon the tool.
Reliability handling around the AI service
The backend discovers which models are currently available, respects rate limits, and retries briefly on transient failures before reporting a clear error. Models get retired, so this is normal operating behaviour rather than an edge case.
Secrets stay server side
The AI key lives in a server function and never reaches the browser, so it cannot be extracted from the page.
Sharing without an account
Public result links let someone see the quality of the output before signing up, which makes the tool easier to recommend to a colleague.
Stack & Rationale
React with Vite
A fast single page app for a tool that is used in short, frequent sessions.
TypeScript
Fewer mistakes in the parts that handle user input and AI responses.
Supabase
Accounts, database and per user data isolation in one platform.
Groq running Llama models
Fast generation, including reading text from screenshots.
Serverless function for AI calls
Keeps the AI key on the server and centralises retry logic.
Recharts
Analytics views without a heavy charting library.
Tailwind CSS and shadcn/ui
A consistent interface that is quick to extend.
Lessons & Outcomes
- •Platform conventions are unwritten rules. Following them is what makes a comment believable.
- •People want options rather than one perfect answer, because the right reply depends on the conversation.
- •Features that remove a login step get used far more than expected.