CodeTonight is a product company. We build the AI substrate — and the provable products that run on top of it: tamper-evident legal AI, and finance records you can verify yourself.
Most AI gives you an answer and asks you to trust it. Ours signs its work, so anyone can check it later — independently, offline, without taking our word for it.
The substrate
Before a product can prove anything, something underneath has to do the reasoning, route the models, and keep the rules. That layer is ours — three pieces that work together.
An AI workhorse built on Claude Code. It plans, writes, tests, and reviews its own work in loops — and keeps going while you sleep, with mechanical safety rails it cannot talk its way around.
Plain version: a tireless senior engineer that checks its own homework and never forgets what it learned yesterday.
One doorway to every AI model — Claude, local on-device models, and others — behind a single interface. It handles rate limits, budgets, and failover, so the work never stops because one provider hiccuped.
Plain version: a universal adapter for AI. Swap the model behind it and nothing upstream has to change.
An open protocol spec that lets different AI systems pass reasoning state to each other in a format they all understand — the wiring that makes the rest interoperable instead of locked in.
Plain version: a common file format so two AI systems can actually hand work back and forth.
The products
Two products built on the substrate, in domains where "just trust me" is not good enough — the law, and money.
DONNA reads legal documents, produces an opinion and the clarifying questions a good associate would ask — and signs every step into a tamper-evident chain. If anyone alters the record later, the signatures break and the change shows.
TL;DR — a legal assistant that keeps the receipts, cryptographically.
Financial records, signed and hashed into a chain you can replay and check independently — on a public testnet today. The point isn't to ask you to believe the numbers; it's to let you falsify them if they're wrong.
TL;DR — "don't trust, verify" applied to a ledger.
Testnet today — a working proof, not yet a production mainnet deployment.
Real-time meeting capture, transcription, and automatic summaries — the same substrate, pointed at the conversations where decisions actually get made, so nothing important falls through the cracks.
TL;DR — the notes write themselves, the actions don't get lost.
How the proof works
The same four-step shape underpins every provable product we ship. No magic — just cryptography applied where it actually matters.
The AI reasons over your documents or data and produces a real result.
Every step is signed and hashed into a chain, linked to the one before it.
The chain is anchored so a later edit can't quietly rewrite history.
Anyone can re-check the signatures offline. If it was tampered with, it shows.
Who builds this
CodeTonight (Pty) Ltd is a South African software and AI company. Two decades of shipping for others — now building our own.
Laurie Scheepers
Founder & Technical Director
Architects the substrate and the provability stack — GRIP, HAL, DONNA, and Falsifiable Finance.
Mia Roos
Executive Director: Business
Leads strategy, partnerships, and the business side of bringing provable AI to market.
Two decades of shipping · selected work
Talk to us
Whether you want provable AI for your own domain, a product trial, or just a straight conversation about what's real and what isn't — we read every message.
development@codetonight.co.za