Two years before we told anyone
E1pi ran inside our own group for two years before we released it. Global Power Source is a battery business in Singapore with fifteen companies across several countries, and between them they run Tally Prime, QuickBooks and SAP. One E1pi reads all of them, in one view.
In those two years one question came up more than any other, first from our own team and then from every owner we showed it to: where does my data go? This post is the long answer. It is also the reason E1pi is built the way it is.
Self-hosted means it lives in your office
E1pi is an on-premise AI server. It installs on any recent Mac or Windows computer in your office, ideally on the same network as your accounting software, and your team reaches it from a browser, an iPhone or a Mac, over the LAN or the internet.
The install is the product. You download the macOS or Windows app and sign in with the account you created on e1pi.com. That is all you do. The AI checks whether Node.js is installed and installs it if not, downloads the E1pi repository from our server, installs every dependency, starts the server, and shows you every command as it runs. Updates arrive the same way, when you choose.
It is the same E1pi we use ourselves every day, not a lite edition or a demo build. We are not open source, but we are not a black box either: the code sits readable on your disk, your IT can read every line, and your admin can see every call the AI makes.
Your data, and your Microsoft 365 tokens, stay home
Everything stays on your premises: the books, the documents, the conversations and the full history of every tool call. Nothing sits on E1pi's servers.
This matters more than it sounds. Business plans from the frontier AI vendors need your files saved on their servers, and that includes Microsoft 365 access tokens with full read access to your mail and documents. With E1pi, search and tool execution run locally on your machine. Only minimal chunks of text, around 500 words, go to the model. Never the entire file. Never a token.
| What | Where it lives | Who sees it |
|---|---|---|
| Books and documents | Your office machine | Your team, by role |
| Conversations and tool calls | Your office machine | Admin, one timeline per user |
| Microsoft 365 tokens | Your office machine | Never sent to the model |
| Text sent to the AI model | Chunks of about 500 words | Only what the task needs |
Access is by passkey with end-to-end device encryption. On the office network that means Face ID or a fingerprint, no passwords. From the internet it means biometrics plus a device certificate issued by E1pi's own CA server, with a schedule window you switch on and off. For companies that want nothing to leave at all, the AI model itself can be installed on your server.
A human holds the pen
We do not run 100% AI processes, and we do not run agent swarms. After two years of daily use we know that even the best frontier models make silly mistakes, and silly mistakes are the hardest to trace. So people start the work, steer it and approve it.
In practice this looks like a compose sheet. When you ask E1pi to reply to a customer, a pre-filled email opens in the app with the To, Cc, Subject and body already drafted. You read it, edit it, and press Send or Cancel. The same sheet exists for Teams and WhatsApp. The AI never sends on its own.
- Payment approval before any payment voucher is created.
- Document approval before anything goes to the freight forwarder or the warehouse.
- Sign-off before a GST return is filed.
- Multi-level approvals on anything, from a holiday application to a three-person approval before a bank transfer above a value you set.
It is not AI versus people. It is your people with E1pi versus your competitor's people without it.
Charged by the work done, not tokens
One fixed monthly subscription per company. No tokens, no credits, no caps, no per-seat fees. We charge for the work done, never for the tokens it took.
A fixed price is possible because of two decisions. First, E1pi hosts its own business model on its own hardware: a LoRA fine-tuned Kimi 2.7, post-trained to work with the E1pi harness and kept current on duty, tax and trade notifications from the authoritative government sources. Second, the harness is built to save tokens. Pipe HTML, for example, lets the AI stream a finished report straight to your browser instead of describing it in prose. Those savings are passed on as the fixed price. Frontier models (Claude, GPT, Grok and others) remain switchable per task when a job calls for one.
Included: all twelve reports, every checklist skill, every trade document, the carrier and packing skills, self-install and continuous updates. The trial runs on your live data, you can cancel anytime, and there is no upfront licence, no year-long contract and no migration project. Talk to us for the figure for your size of business.
The books you already keep, live
E1pi does not migrate your data and does not sync it. It reads live from the ERP you already run, through what we call Data Aware Integration: the model is post-trained to understand each ERP's data format, which data is saved where, and for a given task which data is correct and which is not. Native today: Tally Prime, SAP and QuickBooks. Zoho and Xero are next. Custom ERP and SQL systems are supported the same way.
On top of that integration sit the twelve fundamental reports every company and every auditor needs, in an identical format whichever ERP you run, live and drillable to the ledger and the voucher, for any span, with all your companies side by side for one period.
- Profit & Loss, with GP% and NP% per period
- Balance Sheet, as at any date
- Trial Balance
- Day Book
- Stock Summary
- Bills Receivable, with ageing
- Bills Payable, with one-tap payment advice
- Sales Report
- Purchase Report, including shipping documents and permits
- GST / Tax Report, official exchange rate per row
- Sales Expense with GP, true gross profit per invoice
- Vendor Report
Three ways to run it
We call them Channels, and the point is that your team does not have to learn a new place to work.
- The E1pi app. Web browser, iOS and macOS, with Android coming. Reports, approvals, tasks and conversations, and approvals straight from the lock screen.
- Microsoft Teams. Operate E1pi and assign tasks from the group chat you already live in: reports, approvals and notifications in the chat, a daily digest of due tasks, and a status card after the work is done.
- WhatsApp. Kept for suppliers and vendors. They send bills as a PDF, an image or a handwritten note; E1pi reads each one, checks it against the contract and the quote, routes it for approval, and acknowledges.
Where this goes next
Once AI is inside the workflow, the request that used to need forms, emails and a specialist becomes one sentence, reviewed and approved by a person. The chain we run today: an RFQ email becomes a quote with data sheets, sent as a reply in the customer's own thread. The customer's PO is checked line by line against that quote. One PO becomes the full trade document set. A 3D packing plan confirms the load fits a 20 ft container before anything is booked. The sailing is chosen from live schedules and the documents are uploaded through the carrier's API (booking write access is being enabled by Maersk). The container is tracked to arrival, and the carrier and warehouse bills are checked against the plan before anyone pays.
Every step drafted by AI. Every step approved by a person. Every step announced on Teams or WhatsApp.
Run the business. Let E1pi run the back office. Read more at e1pi.com or open the app at ai.e1pi.com.