Mirror

A vision-model personality identikit. Upload a photo of your bookshelf, desk, or workspace; get back a playful reading of who you might be based on what's visible.

Live demo: https://mirror.psyrcuit.com

What it does

You drag in a photo. A vision model (Claude Opus 4.7, with GPT-4o as fallback) examines the visible objects — books, desk items, posters, the general aesthetic — and returns:

• 5 personality traits inferred from the scene
• A 3-4 sentence synthesis weaving them together
• A one-line "vibe" summary

It's entertainment, not assessment. The output is meant to be playful and shareable, not diagnostic.

Why this exists

Built as Build 1 of 7 in a 90-day public AI demo portfolio focused on AI legibility and personal-corpus tooling. Mirror is the cheap, shareable opener — it demonstrates vision-API competence in something other than the usual "object detection" demo, and the output is screenshot-bait by design.

How to run it locally

git clone https://github.com/psyrcuit/mirror
cd mirror
npm install
cp .dev.vars.example .dev.vars
# Add your ANTHROPIC_API_KEY and OPENAI_API_KEY to .dev.vars
npx wrangler pages dev .
# Open http://localhost:8788

You'll need:

• Node.js 20+
• An Anthropic API key (for the primary vision call)
• An OpenAI API key (for the fallback path)
• Wrangler CLI (npm install -g wrangler, or use npx)

Architecture (one-line summary)

Single static HTML page served from Cloudflare Pages. Browser uploads image, hits a Pages Function, function calls vision API with a structured-output personality prompt, returns JSON, browser renders the card. Nothing stored server-side. Two-layer rate limiting (per-IP + global daily cap) via Cloudflare Workers KV.

Honest limitations

• It's not actually reading your personality. It's reading the scene and producing a playful inference. Don't make life decisions based on it.
• Output quality varies wildly with image quality. Blurry, empty, or person-dominated photos produce weak readings. Bookshelf-with-many-books works best.
• Anglocentric bias. The vision model recognizes Western book covers and aesthetic markers more reliably than non-Western ones. Working on it.
• Single-language output. English only in v1.
• Rate limited per IP. 5 requests per IP per hour. If the demo is getting hammered, you'll see a 429.
• Globally rate limited. Total 40 requests/day across all visitors. If the daily cap is hit, the demo returns a friendly 503 until midnight UTC. Cost discipline > availability for a demo.
• No image storage. This is intentional (privacy), but it means there's no "share this specific reading" URL — you have to screenshot/PNG-export the card to share.
• Costs money to run. Each call hits the Anthropic API. Self-hosting will burn through your API credits if you make Mirror public without rate limits in place. The daily cap protects you from runaway spend; don't disable it without thinking through the consequences.

Privacy

• Your image is sent to Anthropic's API (or OpenAI's, on fallback). Subject to those providers' data policies.
• Nothing is stored on the Mirror server. No database, no logs of image content.
• Cloudflare's standard request logs may temporarily retain IP + timestamp for rate limiting purposes.
• The IP-based rate limiter uses Cloudflare Workers KV with a 1-hour TTL. The daily cap counter has a 24-hour TTL.
• IPs are stored hashed (not raw) in the KV namespace to reduce identifiability of the audit trail.

Self-hosting cost (if you fork and deploy your own)

• Cloudflare Pages: free (unlimited bandwidth, unlimited static-asset requests)
• Cloudflare Workers KV: free at this scale (well under the 100k reads/day, 1k writes/day free tier)
• Custom subdomain via Cloudflare DNS: free if you already have a domain on Cloudflare
• API spend: bounded by your daily cap configuration. Default is 40 Opus 4.7 calls/day at ~$0.024/call = $0.96/day max. Adjust in functions/api/analyze.js if you want a different ceiling.

Total fixed cost: $0/month plus your API spend.

License

MIT. Build derivative versions, fork, run your own. Pull requests welcome but no commitment to merge.

What's next

Possible v2 directions (not committed):

• Side-by-side comparison mode (your bookshelf vs. someone else's)
• Custom prompts ("make it brutally honest" / "make it generous")
• Multi-image upload (multiple angles of the same room)
• Self-hosted OSS deployment guide
• Cloudflare AI Gateway integration for additional spend visibility

If any of these would be useful to you, open an issue and tell me why — that's how the priorities get set.

---

Built with Claude Code. Part of Chase Lance's 90-day AI demo portfolio.