labs@atheron:~/work$ git log --products
Products we run
Every case study is something we built and still run.
We do not show logos of clients we cannot name. Here is what we have built for ourselves and still run, with the live build status straight from our own tracker.
// case studies
Built, tested and in production.
Each one says what we built, with what, and a few facts you can check. The status lines are read live from our public progress tracker.
// 01 web app, members portal, store
ATON Trading
The public site for an AI trading platform, the members portal behind it, and the pre-order store customers buy from.
// stack
- Next.js
- React
- TypeScript
- PostgreSQL
- Stripe
- Tailwind CSS
● live
// what we built
- A marketing site with a pre-order waitlist, built page by page with live market charts drawn in the browser.
- A members portal: plans, orders, gifts and credit, exchange keys, community standing and referrals.
- A pre-order store taking cards through Stripe and Interac e-Transfer, with tax worked out per country.
- One account across the site, the portal and the desktop app, through our own sign-in service.
// facts
- The site's database login can add a waitlist address and do nothing else, not even read one back.
- A test fails the build if any page makes a claim about trading performance.
- Trading is paper only until launch: no customer money is traded before then.
// 02 public blockchain
The ATON blockchain
A public blockchain for payments, specified before it is built, with its own site and a live roadmap.
// stack
- Specification first
- Next.js
- TypeScript
- Canvas
● live
// what we built
- The consensus and protocol specification the node implements.
- The public site, with every claim checked against the specification by a test.
- A roadmap read live from our progress tracker.
// facts
- The chain is a block-DAG, ordered by a GHOSTDAG-style rule: blocks may be made in parallel, and every node holding the same blocks reaches the same order.
- One billion ATON is a hard cap: never more than one billion exist.
- A fair launch: no premine and no sale. Units are issued by mining, apart from a team allocation whose size is disclosed before the first block.
- One memory-hard proof-of-work algorithm, so ordinary graphics cards are competitive.
- What is not decided yet, such as the hash function, is written down as not decided.
// 03 saas live chat
ATON Desk
Live chat for any website, with an AI agent trained on your content answering first and your team second. Coming soon. Ticketing and a knowledge base will be add-ons.
// stack
- Next.js
- TypeScript
- PostgreSQL
- Claude
- Discord
- Telegram
// what we built
- A live chat widget and an inbox for the people who answer.
- An AI agent, running on Claude, trained by scanning your website and from the files and questions and answers you give it.
- Live translation, so a customer and an agent can each write in their own language.
- Discord and Telegram as channels, coming soon.
- Ticketing and a help centre knowledge base as add-ons.
// facts
- Designed for our own support first: Atheron is its first customer.
- When it opens, plans start at $19 CAD per seat a month, with a free plan and a 14-day trial.
- Data is stored in Canada.
// 04 discord and telegram bots
The ATON community bots
Bots for Discord and Telegram sharing one moderation core, run from our team portal.
// stack
- TypeScript
- Node.js
- Discord gateway
- Telegram Bot API
- PostgreSQL
- fast-check
● live
// what we built
- Server and group management, automatic moderation, verification, roles, cases and appeals.
- Scam, impersonation and lookalike detection, and captchas at the door.
- Raids handled as events: moderators are told, every arrival meets a captcha, and it ends by itself.
- Standing, points and a rewards catalogue, worded the same in both rooms.
// facts
- No bot framework: the Discord gateway client and the Telegram API calls are our own.
- The rules live in one shared core, so the two platforms never disagree.
- The bots sign in to the database as their own login and cannot reach pay or applicants.
- A member's standing is asked of the portal, so there is one definition of the score.
// 05 back office, customer portal, sign-in
The members and team portals
Our customer portal, the back office our team runs the company from, and the sign-in service both use.
// stack
- Next.js
- React
- TypeScript
- PostgreSQL
- Stripe
- Microsoft Graph
● live
// what we built
- A team portal with HR, Finance, CRM and Developer areas, gated role by role.
- A members portal with plans, orders, a pre-order store and community pages.
- Our own sign-in service: an OpenID Connect provider with passkeys, authenticator codes and recovery.
- Wallet sign-in for EVM, Bitcoin, Solana, TON, Sui, Aptos and Cosmos wallets, with the challenge made on the server.
// facts
- Rules that matter are kept in the database, so they hold whichever app is talking to it.
- Each service signs in to the database as its own login, allowed only what it needs.
- Servers are snapshotted and databases dumped every eight hours.
- A privacy inventory says how long every column of personal data is kept, and a test fails when one is missing.
// 06 ai models and agents
ATON AI and our agentic tooling
A trading model we train on our own GPUs, a news labeller running on an open-weight model, and the automated checks that review our code.
// stack
- XGBoost
- PyTorch
- CUDA
- Hugging Face transformers
- Claude
- Python
- TypeScript
// what we built
- ATON AI: gradient-boosted trees trained on the GPU, with a neural network that is used only if it beats them on the same held-out data.
- Continual training: one exchange at a time, every interval from one minute to one week, with the latest six months held out.
- A backtester that replays our live rules on the GPU, queued from the team portal.
- A news labeller running an open-weight model locally, tagging events, direction, urgency and topic, with no hosted API.
- A verification server that re-runs our tests, part of the mutation set and secret scans every four hours, and reports what it finds.
// facts
- Trained on our own desk GPUs, with no cloud GPUs: one card for the trading model, one for the language model.
- A governor holds the main card's load down while someone is working at the desk, and never leaves a job suspended.
- Every read of the held-out months is logged, so the backtests stay honest.
- The team portal's assistant runs on Claude and has no tools by design: it reads, drafts and summarises.
// breakouts
Every project, broken out.
What each one is, how it was built, the size of it, and what a comparable build would be quoted at today. The price is worked out from our rate card while you read it.
// start
Want something built like this?
The same team, the same tests, the same care. Tell us what you need.