Cherry 🍒

Name any product. Cherry reads what the internet is saying about it, picks the few issues worth acting on, and tells you what to do — grounded in real sources.

Studio Felix · built on the Anthropic API.

A live, AI-native customer-feedback triage tool — a working v1 of an operating system for voice-of-the-customer. It intakes signal two ways (search the open web, or paste raw internal feedback — a Slack thread, support tickets, sales-call notes), synthesizes the noise into ranked, source-grounded issues, routes each to the team that owns it, and closes the loop with drafts and status — with a human in the loop at every step and evals measuring quality.

• Intake, two ways — search the public web, or paste internal feedback and Cherry triages it directly (no web search). The signal lives inside your company too.
• Legible signal — each issue scores severity, reach, and recency as separate 1-5 axes; a live Signal = severity·reach·recency control lets you re-weight what "high-signal" means and the ranking re-sorts in place. No opaque number.
• Screens out fake signal — public reviews get gamed (bought 5-stars, planted 1-stars, review-farm text). Cherry scores each issue's authenticity 1-5, weights verified-provenance sources (App Store, G2) over anonymous ones, collapses templated bot clusters, and flags issues that may not come from real users — so astroturf doesn't inflate your priorities.
• Human-in-the-loop & self-improving — disagree with a call, tell Cherry what's off, and it re-ranks with your judgment as ground truth. Corrections persist and ride along into future queries; an on-page metric shows your correction rate dropping.
• Routes the work — a one-line digest (3 Engineering · 1 Billing · 1 Legal) reframes the list into who owns what; each issue's ticket is framed for that team, and tickets can be sent straight to Slack.
• Closes the loop — every issue has a lifecycle (new → triaged → routed → shipped) and Draft ticket / reply / "you said, we did" update, so Cherry does the first-pass writing and you can trace signal to outcome.

How it's wired

  browser (index.html)  ──POST {name, corrections}──▶  /api/triage  (serverless)
                                                          │ holds the API key
                                                          │ Claude + web search
                                                          │ → ranked, cited JSON
  ◀────────────────── clean result JSON ──────────────────┘

The API key lives only in the backend as an environment variable — never in the browser, never in the repo. Corrections persist client-side (localStorage) and are sent up as learned context, so there's no database to provision. The model's answer is schema-enforced (Anthropic structured outputs, output_config.format) so the response is always valid JSON in the triage shape — no parsing free text. Each issue must carry a real source URL or it's dropped, and an automated eval gates every change in CI (evals/check.mjs).

Run it locally

cp .env.example .env.local      # put your real ANTHROPIC_API_KEY in it
npm i -g vercel                 # one-time
vercel dev                      # serves index.html + /api/triage on localhost

Open the local URL, type "Notion", and you'll get a live, cited result.

Deploy it (Vercel — ~5 clicks)

1. Push these files to a GitHub repo (e.g. stephiesworld/cherry) — index.html at the root, with the api/ folder alongside.
2. Go to vercel.com → sign in with GitHub → Add New… → Project → import the repo.
3. Framework preset: Other (no build step — it's static + one function).
4. Environment Variables → add ANTHROPIC_API_KEY = your key. Deploy.
5. You get a public URL like cherry-xxxx.vercel.app. Type a product to test.
6. (optional, recommended) add CHERRY_ALLOW_ORIGIN=https://<your-url> and redeploy to lock the backend to your own site.

Every push to the repo auto-redeploys.

Config (all optional — defaults are sensible)

Env var	Default	What it does
ANTHROPIC_API_KEY	—	Required. Your key; backend-only.
CHERRY_MODEL	claude-sonnet-4-6	Fits the free-tier 60s limit. Set claude-opus-4-8 on Vercel Pro.
CHERRY_MAX_SEARCHES	3	Web searches per query (the main cost/time lever). 3 keeps comfortable headroom under the free-tier 60s limit; raise on Vercel Pro.
CHERRY_DAILY_CAP	200	Hard ceiling on triage queries/day (abuse guard).
CHERRY_DRAFT_DAILY_CAP	300	Hard ceiling on draft requests/day.
CHERRY_PER_IP_PER_MIN	6	Per-visitor rate limit.
CHERRY_ALLOW_ORIGIN	*	Lock CORS to your site in production.
CHERRY_SLACK_WEBHOOK	—	Slack Incoming Webhook URL; enables "Send to Slack" routing.

Cost & safety

Roughly $0.15–0.20 per query (the web search is most of it; the model adds ~5¢). A handful of users clicking around ≈ a few dollars. The guards that keep it there: identical product names are cached 24h (repeat plays are free), a per-IP rate limit, and a hard daily cap. As a backstop, set a spend limit on your key in the Anthropic Console. The in-memory cache/limiter is per-instance — fine for a portfolio tool; back it with Vercel KV for real scale.

Quality gate (evals)

Two layers — structure and quality:

npm run eval          # check.mjs  — structural contract (deterministic, runs in CI)
npm run judge         # judge.mjs  — LLM-as-judge synthesis quality (needs ANTHROPIC_API_KEY)

check.mjs treats the triage output as a product with a contract and fails loudly when it's violated: every issue grounded in a real source, severities in range, the list actually ranked, every action justified. It runs automatically in CI on every push (.github/workflows/eval.yml), so a change that breaks the contract fails the build before it ships — closed-loop quality, not vibes.

judge.mjs goes further: a second Claude call grades synthesis quality — clustering, ranking defensibility, routing correctness, grounding, actionability — each 1-5 against a rubric, and appends the score to judge-history.jsonl so the trend is visible as the prompt evolves. That's the harder eval problem: catching a triage that's structurally valid but worse. (golden.json is an illustrative recorded result; pipe a live one in with curl … | node evals/check.mjs -.)

Files

Path	What
index.html	The front end (Studio Felix design; calls its own backend).
api/triage.js	Triage backend: web-search or pasted-feedback intake, structured-output JSON, memory, guards.
api/draft.js	Drafts a ticket (framed for the owning team), customer reply, or "you said, we did" update.
api/route.js	Routes a drafted ticket to Slack (CHERRY_SLACK_WEBHOOK).
evals/check.mjs + golden.json	Structural quality gate (CI).
evals/judge.mjs	LLM-as-judge synthesis-quality eval.
DESIGN.md	The brand system.

— Built with Claude. Studio Felix.