MULTIVAC

A Knowledge Vending Machine: Retrofitting Legacy Hardware with AI-Powered Information

Project Vision

MULTIVAC is a vintage receipt-printing vending machine retrofitted with a Raspberry Pi and AI. You insert payment, ask a question, and get an answer printed on a receipt. It's like a payphone for knowledge—imagining a world without smartphones.

The machine uses a MEMS microphone for voice input, an RFID reader for payment, and a custom PCB to control the original thermal printer. Queries go through OpenAI's API, and responses come back as printed receipts—making information tangible.

Ideation & Iteration

Concept A: Smart Mirror

REJECTED

Initial idea: smart mirror with voice commands. Too ephemeral—information on a screen doesn't feel real. I wanted something physical.

Concept B: Kiosk

EVOLVED

Wooden kiosk with e-ink display. Too bulky and slow. But the concept of a physical device sparked the next idea.

Final: MULTIVAC

SELECTED

Found a vintage receipt printer at e-waste. The thermal printer creates a beautiful physical artifact. Modern guts, retro exterior.

Development Phases

Phase 1: Disassembly

EXPLORATION

Took apart the machine to understand the thermal printer circuitry, card reader, original LCD, and power system. Documented everything to plan the retrofit.

Phase 2: Custom Printer Control

PCB DESIGN

Designed and milled a custom PCB to control the thermal printer from the Pi. Handled power distribution, serial communication, and heat management.

Phase 3: Audio Input

INPUT SYSTEMS

Added a MEMS microphone with audio filtering and voice activity detection. Processes speech for OpenAI's Whisper API.

Phase 4: Payment Integration

PAYMENT SYSTEMS

Integrated an RFID reader for payment. Implemented transaction handling and balance tracking while keeping the original housing intact.

Phase 5: Display & UI

OUTPUT & FEEDBACK

Installed a modern LCD screen. Built a simple interface showing recording status, payment, and confirmation before printing.

Phase 6: Raspberry Pi Setup

CORE COMPUTATION

Installed a Pi Zero W inside the original housing. Configured WiFi for reliable network access in public spaces.

Phase 7: Software

INTELLIGENCE

Built Python control system. Integrated OpenAI's Whisper (speech-to-text) and GPT (responses). Handles all hardware coordination.

Python OpenAI API Whisper & GPT

Phase 8: Housing & Aesthetic Retrofit

PHASE: INDUSTRIAL DESIGN

Critical design challenge: maintain visual similarity to the original while accommodating modern components. Used selective 3D printing, careful wiring routing, and subtle modifications to housing. Preserved branding and physical character of the original device.

3D Printing Industrial Design Fabrication

Final: Testing, Refinement & Deployment

PHASE: VALIDATION

Comprehensive system testing: voice recognition accuracy, printer reliability under continuous use, payment security, network stability, and user experience flow. Iterative refinement based on real-world usage patterns. Device ready for public deployment.

QA Testing User Testing Deployment

Project Documentation & Media

Photos, videos, design files, CAD models, and manufacturing documentation

Functional demonstration: Device activated by RFID signal, processing queries and printing responses
Project summary slide
Project summary slide
For this project I wanted to retrofit an old device with cutting edge software and leave it looking as inconspicuous as possible, like the trend of sleeper build computers. I want there to be an element of magic or fantasy in an old device inexplicably having modern capabilities. That's why, when I found this old ESC/POS point of sale machine in a mail room trash center, I knew it would be perfect for my work.
For this project I wanted to retrofit an old device with cutting edge software and leave it looking as inconspicuous as possible, like the trend of sleeper build computers. I want there to be an element of magic or fantasy in an old device inexplicably having modern capabilities. That's why, when I found this old ESC/POS point of sale machine in a mail room trash center, I knew it would be perfect for my work.
Testing my LCD screen.
Testing my LCD screen.
Soldering this MEMS microphone was by far the most technically challenging part of the build. Here it is compared to my finger.
Soldering this MEMS microphone was by far the most technically challenging part of the build. Here it is compared to my finger.
I have very limited electronics design experience, so this was really my first time working with a buck converter. I did not want to pull the power for all the peripherals from the Pi's onboard 3.3v rail and risk a brown out, so I stepped down the 5v using a buck converter and ran that into the RFID reader, as can be seen in this photo.
I have very limited electronics design experience, so this was really my first time working with a buck converter. I did not want to pull the power for all the peripherals from the Pi's onboard 3.3v rail and risk a brown out, so I stepped down the 5v using a buck converter and ran that into the RFID reader, as can be seen in this photo.
Connecting the pi to the proprietary PCB on the thermal printer was tough. Here is a photo of the size of the leads I had to solder to.
Connecting the pi to the proprietary PCB on the thermal printer was tough. Here is a photo of the size of the leads I had to solder to.
Milling the custom PCB for the internals.
Soldering male headers onto my RFID reader.
Soldering male headers onto my RFID reader.
Soldering the pi GPIO pins.
Soldering the pi GPIO pins.
I started by prototyping with direct jumper wires before soldering in case any issues arose. Here is the first photo of my fully functional prototype. pictured is the Pi, LCD screen, RFID reader, microphone, and printer.
I started by prototyping with direct jumper wires before soldering in case any issues arose. Here is the first photo of my fully functional prototype. pictured is the Pi, LCD screen, RFID reader, microphone, and printer.
The first build looks almost identical to a normal point of sale machine!
The first build looks almost identical to a normal point of sale machine!
Testing output.
Testing output.
Testing again.
Testing again.
Here is where I found the machine. As you can see, I have retrofitted it to look as similar as possible to the original ones.
Here is where I found the machine. As you can see, I have retrofitted it to look as similar as possible to the original ones.
Gerber file for the custom PCB I designed for the device.
Gerber file for the custom PCB I designed for the device.
KiCad screenshot with custom PCB for 40 pin Raspberry Pi Zero W, following design rules for Arch Shop PCB mill.
KiCad screenshot with custom PCB for 40 pin Raspberry Pi Zero W, following design rules for Arch Shop PCB mill.

Design Files & Resources

KiCad project files, Gerber outputs, 3D housing, and firmware code

project_reduced-F_Cu.gbr

Gerber file for PCB copper layer

⬇ Download
project_reduced.kicad_pcb

KiCad PCB layout file

⬇ Download
housing3x5.stl

3D printable custom PCB housing (78 MB)

⬇ Download
fullpipeline.py

Complete firmware code for device programming

⬇ Download

Device Firmware

Complete Python implementation for MULTIVAC device control and AI integration

Download full source code →

Project Specifications

What does it do?

A machine that receives payment and produces knowledge in physical form. Like a payphone for information. Imagine a world without smartphones—this would exist in convenience stores and libraries.

What's been done before?

I've seen thermal printer builds, Raspberry Pi + LLM integrations, and DIY payment systems. But I didn't copy any single design—this combines pieces in a new way.

What did I design?

Electronics, housing, PCB, information pipeline, GUI, and the core idea. Everything from hardware integration to software architecture.

Materials & Components

Housing: PLA (3D printed)
Printer & case: NuPrint210 (scavenged from Harvard's mailroom)
PCB: Copper (arch lab)
Electronics: Raspberry Pi Zero W ($13), RFID RC522 (archshops), MEMS mic ($6.95)

Tools & Processes

KiCad design, 3D printing, PCB milling, cutting, filing, soldering, and integration.

What worked? What didn't?

Worked: Printer, AI pipeline, voice input, display.
Didn't work: Payment terminal (payment processing is more complex than expected; I have an RFID scanner reading generic cards instead).

Course Requirements Met

Gratitude

This semester has been extraordinary. I want to thank Adin, the architecture TAs Jen and Gert, and Neil for their mentorship and support. The friends in the Architecture Section made this work possible, and I've learned more in these few months than I could have imagined. Thank you for believing in the ideas, providing feedback, and making this an unforgettable experience.