How Kopi works.

Why

Approvals leak at the doc-to-form handoff

Most SMB approval flows live as policy docs nobody reads, plus a Slack / Lark / email chain that re-derives the same rules every time. A vendor gets paid without an invoice. A reimbursement skips the cap. A leave request bypasses HR. The handoff between “what the policy says” and “what the form actually enforces” is where the leakage lives.

Kopi closes that loop. Mirror reads your existing docs, extracts the form + rules + routing, and runs the resulting spec against every submission. When a rule fires she cites the exact sentence from the source. No hallucinated rules, no “trust me, this is policy.”

Get going

Quick start

1. 1

   Sign up

   Visit hellokopi.com/signup — magic link or Google OAuth, no install.

2. 2

   Meet Mirror

   Land on /app. Mirror greets you with three chips: Show me a demo / Drop my docs / I'll explore. Pick demo for an instant working workspace, or skip straight to your real docs.

3. 3

   Drop a policy doc

   PDF, PNG, JPG, or WebP. Up to 5 files, ≤8MB each. Mirror runs Stage 1–5 and surfaces the draft in ~3 minutes.

4. 4

   Review the draft

   Four tabs — Fields, Rules, Routing, Currency & Compliance. Every item has a verbatim quote and a confidence pip. Edit inline or ask Mirror in chat.

5. 5

   Apply & go live

   You get a public form URL — hellokopi.com/forms/<workspace>/<slug>. Share with your team. Submissions auto-clear when they pass; flag yellow when a rule fires; block red on a hard cap.

Architecture

Source → Event → runTemplate → Surface

The conceptual model is four layers. Sources ingest external input (web form, email, Slack, Lark, API) into a durable ApprovalEvent. Surfaces render the event for an actor (HTML email, web review page, Slack DM, Lark card). The decision layer runTemplate between them is pure code, channel-agnostic.

   Source (Web · Email · Slack · Lark)
            │
            ▼
   ApprovalEvent (durable, channel-agnostic record)
            │
            ▼
   runTemplate(spec, request) → Decision      ← pure code, ~4ms, no LLM
            │
            ▼
   Surface (HTML email · /approvals · Slack card · Lark card)

• Spec is data, not code. Mirror generates a TemplateSpec JSON per template; the runtime compiles it at request time. No per-tenant TypeScript.
• Sources ingest, Surfaces render. Web / Email / Slack / Lark are interchangeable on both ends; the decision layer doesn’t care.
• Citation contract is non-negotiable. Every field, rule, and routing approver carries a verbatim quote. No quote → drop the item.

Stack

What it’s built on

Frontend + API

Next.js 14 (App Router) on Vercel Fluid Compute

Database

Supabase (PostgreSQL)

AI

Claude Sonnet 4.6 via Anthropic API

Auth

Magic Link via Resend + Google OAuth

Email

Resend (transactional + inbound)

Storage

Supabase Storage

Analytics

Vercel Analytics + GA4

Channels

Slack OAuth, Lark webhooks, web forms, inbound email

Page map

Where things live

Status

Where we are

• Phase 1–2

  In production

  Lark webhook + 5-dimension AI check + three-color cards + auto-learning rule library + admin dashboard. Live with a paying customer.

• Phase 3

  Active (Q2 2026)

  Multi-channel platform + Mirror. Web + Email is the default; Slack and Lark are accelerators. v1 shipped on hellokopi.com — 5 tenants, 3 active templates, 70+ approval events.

• Iterating

  Now

  Onboarding UX (demo workspace, first-run nudges, four-tab Settings), submission detail UI, Mirror extraction precision (currently 0.89 overall confidence on real customer docs).