Kashify

A personal finance ledger you talk to. Send a WhatsApp message — "lunch 850" — and an AI agent ("Neo") parses it, categorizes it, and your web dashboard updates seconds later. No forms, no app to open. Built for people who juggle two currencies (ARS + USD) and think in Rioplatense Spanish.

Private beta · multi-user with per-user data isolation

🔗 Live demo · 🎬 2-min walkthrough

---

Why

Expense tracking dies on friction. Kashify removes it: the capture surface is the chat app you already have open. Neo turns natural-language messages into structured transactions — the web app is for seeing your numbers, not data entry.

What it does

• WhatsApp capture — natural language → transactions via Neo (Claude). Handles "café 1200", "cobré 50 lucas", ARS/USD, income vs. expense.
• Neo learns — when it doesn't recognize a keyword it asks and remembers the keyword→category mapping (fuzzy-matched with Fuse.js) instead of guessing or burning tokens. A NEO_LLM_FALLBACK=false switch lets Neo run at zero LLM cost.
• Multi-currency ledger — ARS + USD side by side (the Argentine "dollar cushion" is a first-class concept, not an afterthought).
• Web dashboard — balances, charts (Recharts), category breakdowns. PWA-installable.
• Installments (cuotas) — splits a purchase into scheduled monthly entries.
• Budgets & savings goals — period budgets and goal tracking.
• Bulk import — ingests monthly spreadsheets (xlsx/csv), including a category × month matrix that it unfolds into individual movements.
• Persistent appearance — per-user theme + icon prefs stored in the DB.

Architecture

WhatsApp (Meta Cloud API)
        │  webhook
        ▼
Cloudflare Worker ──fast ACK──►  Vercel Cron (signed: CRON_SECRET)
                                        │
                                        ▼
                         Next.js API (App Router)
                          ├─ Neo: parse / learn (Anthropic Claude)
                          └─ Ledger logic (installments, budgets, import)
                                        │
                                        ▼
                         Supabase (Postgres + Row-Level Security)
                          every query scoped to the user

A Cloudflare Worker absorbs the WhatsApp webhook and the heavy work runs off a Vercel Cron, so Meta always gets a fast 200 while parsing happens out of band.

Stack

Next.js 16 · React 19 · TypeScript · Tailwind v4 · Supabase (Postgres/RLS) · Anthropic Claude · Cloudflare Workers · WhatsApp Cloud API · Recharts · Fuse.js · Vercel

Run locally

npm install
cp .env.local.example .env.local    # fill Supabase, Anthropic, WhatsApp keys
# apply supabase/migrations/* to your Supabase project
npm run dev                         # → http://localhost:3000

WhatsApp is optional for local dev — the web app and Neo's parsing work without it. See .env.local.example for the full variable list.

Status

Private beta (allowlisted users). Single-developer project — product, design, stack and infra by Francisco Remonda.

Repo: github.com/franremonda-cmyk/Kashify