Most people open Google Sheets to handle monthly budgets, invoices, or attendance logs. I wanted to see if spreadsheet cells could work as an actual computer chip. That led to SSX Silicon 1, a working 32-bit computer processor built inside a single Google Sheets workbook. It runs with 32 registers, a 16-kilobyte memory space, 36 instructions, and a 32 by 32 pixel color display. The sheet does not use pre-recorded animations, placeholder numbers, or external file links. Every calculation runs through a five-stage pipeline: fetch, decode, execute, memory access, and writeback.

You can run the processor in two ways. With Google Sheets iterative calculation turned on, toggling a manual clock cell advances the machine cycle by cycle using native circular formulas. For continuous execution, an embedded Google Apps Script V8 engine powers interactive buttons and a custom top menu, running programs at hundreds of instructions per second. The 32-bit arithmetic logic unit handles numbers up to 4.29 billion across 36 instructions, covering arithmetic, division, bit shifting, logic gates, and subroutine branching. It also tracks four hardware status flags on every cycle: Zero, Carry, Sign, and Overflow.

The memory layout organizes a self-contained 16-kilobyte address space across dedicated sheets. This includes 4,096 words of program ROM, 4,096 words of read-and-write data RAM, system I/O registers, and a stack region starting at address 0x3FFF. Thirty-two registers handle active computation: R0 stays fixed at zero, R1 through R27 store math and variables, and R28 through R31 serve as pointers for the frame, stack, return address, and program counter. A dedicated video RAM block at 0x2400 maps directly to a 32 by 32 grid of 1,024 cells, using conditional formatting and a 16-color RGB palette to draw borders, crosshairs, and pixel graphics. I also wrote a Python two-pass assembler that turns human-readable .ssx assembly code into spreadsheet-ready CSV tables, along with interactive HTML architecture diagrams built with Archify.

To make sure the CPU calculates accurately, I tested every instruction against a Python reference emulator across five benchmark programs. The hardware self-test verifies all 32 arithmetic operations, memory reads, and stack frames in 77 cycles, leaving the golden signature 0xCAFEBABE in register R1. A 32-bit Fibonacci program generates 47 terms in RAM up to 1,836,311,903 without rounding or overflow. A prime number sieve finds all 25 primes under 100 using memory arrays, an integer matrix program calculates a 3x3 matrix product, and a graphics routine draws a multicolored geometric crest across 232 lit pixels. The project includes five PDF documents: an academic research paper, an instruction set manual, a microarchitecture specification, a verification report, and a beginner guide. Everything runs in a standard web browser. For anyone who wants to open the processor in Google Sheets, check out the Python compiler, or read the manuals (including the step-by-step Beginner Guidebook and Research Paper), the full open-source release is available on my GitHub!

Inside the repo, you will find:

The ready-to-import Google Sheets CPU workbook (.xlsx)

The Python assembler & bit-exact reference emulator

Interactive HTML architecture & pipeline diagrams

All 5 PDF manuals and the 5 sample assembly programs

Your step by step guide are so helpful

AI Customer Support Agent for Web Form, Gmail & WhatsApp An AI support agent that answers customer questions across a web form, Gmail and WhatsApp, and hands off to a human team when it should. Load-tested with 200+ messages and no lost messages.

I built an AI customer support agent as a demo for a fictional SaaS company called Flowmint. Customers can write through a web form, Gmail or WhatsApp. The agent finds the answer in a knowledge base, replies on the same channel, and passes the conversation to a human team (Billing, Sales, Support) for refunds, special pricing, or when a customer asks for a person.

One agent for three channels, with replies formatted for each (a clean HTML email, a short WhatsApp message, a private web tracking page)

Safety rules written in code: it can only quote approved prices, and refunds always go to a human

Remembers returning customers on email and WhatsApp

Private ticket tracking page with a secret tracking code

If the AI service is busy, it switches to backup AI models, and to a knowledge-base answer if all are limited

I load-tested the agent for 24 hours with 200+ messages across all three channels, using simulated customers plus real emails and WhatsApp messages I sent myself.

212 messages in, 212 replies out, none lost

Only 3.3% needed a human

Areas I’m improving: faster replies with a paid AI plan, and spam protection for the web form.

Tech stack: Python, FastAPI, OpenAI Agents SDK, Kafka, PostgreSQL, Next.js, Docker, Gmail API, WhatsApp Cloud API.

Writing the safety rules in code, not in the prompt, is the part I like most: approved prices only, and refunds always go to a human. My RAG demo also passes the question to a person when the guide has no answer. How does your agent decide that a customer "asks for a person": a fixed list of phrases, or does the model decide?