JIFFY | Less Shit. More Life.
Jiffy
Jiffy Multimodal Agent
Capabilities
Checking account balance using text and sending money by voice commands
Parsing Receipt and figuring out participants from text message
owner claim form
Receipt Dashboard
Auto payment requests send over whatsapp
Abhinav's share claim page
Bill Page for Abhinav
login to bunq
pin for bunq login
payment page
Live updates from Jiffy
Settled Dashboard

Jiffy

Less shit. More life.

Jiffy is an agentic payment multimodal that simplifies payments, gets shit done effortlessly and saves time for users to enjoy life. All in a Jiffy!

01 The Problem

Every group payment starts the same way: one person taps their card, covers the bill, and quietly inherits a second job. They must reconstruct who ordered what, calculate individual shares, dispatch separate payment requests, and then wait, follow up, and chase. 61% of adults have covered a group expense expecting repayment, and 59% report a negative outcome — lost money, strained relationships, or worse (CreditCards.com, 2024). And even for a simple one-to-one transfer, the friction is real: opening a banking app, locating a contact, entering an account number, typing an amount, and confirming adds up to five to seven manual steps for a routine €20 payment. The infrastructure exists. The P2P market is heading toward $14.5 billion by 2034 (Allied Market Research, 2024). What has not existed, until now, is a tool that removes the burden entirely. Jiffy does.

02 The Solution

Jiffy is a multimodal payment assistant. Talk to it, show it things, and it handles the rest.

Send money

Say who to pay and how much — by voice, text, or both. Jiffy resolves the recipient, shows a confirmation, and waits for a thumbs up to send via Bunq.

"Send €50 to Barnali" → Recipient confirmed → thumbs up → payment sent

Receive money

Show Jiffy the bill, name who was there. Jiffy reads the receipt and sends each person a personal WhatsApp payment link. A thumbs up dispatches the links. Everyone marks their own items and pays. You track it all on a live dashboard.

"Collect from Abhinav and Barnali" → Receipt scanned → thumbs up → links sent → live dashboard

Each person selects only what they consumed. The payer never calculates anything.

End-to-end flow

01 Capture	02 Parse	03 Distribute	04 Pay	05 Track
Photograph receipt	AI reads line items	WhatsApp links sent	Each person pays own share	Live dashboard updates

WhatsApp automation

No app download, no sign-up. Jiffy sends each person a personalised link via WhatsApp Web (Playwright), using the organiser's existing account. One QR code scan connects the session; it persists from there.

03 How AI Is Used

Each AI component is purpose-built for its task, using the most efficient model at each layer.

Component	Detail
Intent classification — Claude Haiku (`claude-haiku-4-5`)	Classifies messages into 8 intents in real time, extracting names, amounts, and receipt IDs as structured slots. Supports English, Hindi, and Dutch.
Receipt parsing — Claude Sonnet (`claude-sonnet-4-6`)	Claude Vision reads receipt photos and returns line items and prices as structured JSON. No OCR or templates needed.
Participant extraction — Claude Sonnet (`claude-sonnet-4-6`)	Natural language ("Dinner with Priya and Dev") becomes a structured recipient list in one inference pass.
Conversational agent — Redis state machine	Session-aware state machine retains context across turns, inheriting prior slots (e.g. amount) automatically.

04 Multimodal Input

Four input modes — image, voice, text, and gesture — all feeding the same agent pipeline. No keyboard required at any point.

Modality	Implementation & Role
Image	Receipt photo processed by Claude Vision. Returns structured line items instantly. No OCR needed.
Voice — faster-whisper	Hold-to-talk, transcribed locally via faster-whisper (CPU, int8). No external API. English, Hindi, Dutch.
Text	Natural language chat input for sending payments, naming participants, and checking status.
Gesture — MediaPipe Hands	Thumbs up confirms, thumbs down cancels. Triggered at every payment confirmation step across both send and receive flows. Runs in-browser, no hardware needed.

Photograph bill → name participants by voice → thumbs up to confirm. Zero keyboard interaction.

05 Infrastructure

Fully containerised on AWS EC2, deployed as a single Docker Compose stack.

Component	Role
Frontend (Next.js) + Backend (FastAPI)	Separate Docker images, stored in ECR, pulled to EC2 on deploy.
nginx	Reverse proxy with SSL termination. Routes `/api/*` to FastAPI, all else to Next.js.
Redis	Session state, receipt data, and payment tokens. Persisted via Docker volume.
CloudFront CDN	Global distribution and HTTPS at the edge.
Persistent volumes	WhatsApp session, Whisper cache, and Bunq auth survive container restarts.

06 Tech Stack

Layer	Technology
Frontend	Next.js 16, React 19, TypeScript, Tailwind CSS 4
Backend	FastAPI, Python 3.11
Storage	Redis
AI	Claude Sonnet (receipt scan, participant extraction), Claude Haiku (intent), faster-whisper (voice), MediaPipe (gesture)
Payments	Bunq API (sandbox)
Messaging	WhatsApp Web via Playwright
Infra	Docker Compose, nginx, AWS EC2 + ECR + CloudFront