We built JY Global's internal AI portal: a personal assistant for reminders, tasks, the CRM, invoices and payments; a knowledge hub that answers with citations; a CRM the AI runs, with forms clients fill by link; a chat workspace for the team; and a WhatsApp bot that reads client payments and syncs them to Xero. The AI proposes and a person confirms, every time.
Proposal to production
22 daysfrom the proposal to the live portal00Project overview
JY Global advises CFD brokers and financial-services operators on compliance and governance across Cayman, Mauritius, Cyprus, the UK, Australia, Hong Kong and Malaysia. Codech took the firm from a proposal to a live AI portal in 22 days, then kept shipping: 319 pull requests by the end of September.
The challenge
Every client carries its own licence renewals, AML refreshes, filings and reporting cycles. Across the portfolio that meant hundreds of dates in one legacy spreadsheet and in senior staff's memory, the same memos rewritten by hand, and two advisers able to answer the same question differently.
Our proposal
We proposed the portal as separate fixed-fee modules, locked a design system taken from the firm's own website, and built a 30-page clickable prototype the Managing Director reviewed before any production code. Every production screen is a faithful port of it.
The solution
One portal with five modules: a personal assistant, a knowledge hub, an AI-run CRM with client forms, a chat workspace and WhatsApp payments. The AI drafts, reads and suggests; nothing is written until a person confirms it.
01The problem
Licence renewals, AML refreshes and filings for every client, tracked in one spreadsheet and in people's memory. Reminders had to reach staff wherever they were.
Jurisdictional nuance lived with whoever had handled it last, so the same question could get two different answers. The firm needed one source that cites itself.
The firm gives advisory work only. The AI could draft, read and suggest, but it could never write anything a person hadn't approved.
02Personal Assistant & payments Live
A tool-calling assistant with memory across sessions. It answers questions about the whole client portfolio, sets reminders and tasks for you or a teammate, works the CRM, and keeps track of invoices and payments, including payments clients post in their WhatsApp groups. Every action arrives as an editable card that a person confirms.
Portfolio questions
The assistant reads the client register and key dates, so "what's due for Mauritius clients in the next two weeks" gets named deadlines in order of urgency, plus an offer to draft the follow-ups.
Reminders & assignments
Ask in plain words and the assistant fills in a reminder card: who, when, which client and which channels. You can edit any field before confirming. Reminders fire in the app, as a phone notification and on Telegram.
WhatsApp payments
A listen-only bot sits in each client's WhatsApp group. When a client posts a payment and its slip, the AI matches it to the right invoice and asks a person to confirm. The confirmed payment is marked paid in the portal and in Xero.
Documents
The assistant drafts memos and follow-ups in the firm's format and generates them as Word, PDF, Excel or PowerPoint files. The draft is shown first, and the file is only created once you approve it.
03Knowledge Hub & Chat Workspace Live
The AI learns the firm's procedures, policies and templates and answers with citations, or says it doesn't know. In the chat workspace, staff work on shared projects with the AI reading every document in the project.
Cited answers
Questions are answered by retrieval over the firm's own entries, with the passages it used cited inline. Each entry shows how often it has been cited, its version and when it is next due for review.
Chat Workspace
Staff collaborate in project chats with the relevant documents in context. The AI reads them, answers with section references, and helps the team split up the work, so everyone stays on the same page.
Honest gaps
When nothing in the library covers a question, the hub says so and offers to flag the gap for an owner, rather than filling it from the model's general knowledge. For a firm that gives no legal opinions, that matters more than a fluent answer.
Publishing
Scanned PDFs are read with OCR, and tables in Word and Excel files are extracted. Before a new entry is published, the hub checks it against the library and warns when it overlaps an existing one.
04AI CRM & forms Live
Record types, fields and tabs are fully customisable. The AI can retrieve anything and perform any action on a record, always through a confirmation card. Application forms are drafted with AI, sent by secure link, and answers land on the client's record.
Client forms
When a client uploads a document to their form, the AI reads it and pre-fills the answers it found. Nothing is saved until the client confirms, and staff then verify or query each answer before it reaches the CRM record.
AI-run CRM
Each client record opens on an AI brief built from approved registry facts and linked work, with each claim linked to its source. The record's copilot can propose a change, such as a new key date, and shows it on a card for you to confirm.
Entity builder
Record types, their fields, templates and tabs are data the owner edits, not code. A new type, such as Fund, gets its own register, and changes wait as drafts until they are published.
Access is data too: owners edit a permission grid per role, record groups scope which clients each member sees, and the AI can only do what the person's role allows elsewhere.
The key design decision
The AI proposes. A person confirms.
JY Global gives advisory work only, never legal opinions, so the product couldn't either. The AI reads, drafts and suggests, but it has no way of writing to the database on its own.
Every write it wants to make arrives as an editable confirmation card. A person checks it, and only then does an ordinary API route perform the write, with the same permission check as any other screen and an audit record of who approved it.
05How it works
Whether the request comes from the chat, Telegram, a client form or a record, it goes through the same path. All model calls run through one server-side orchestrator; the browser never calls the model.
06Integrations
Staff work in the portal, on their phones and in Telegram. Everything runs on one stack the firm owns.
Channels02
AI01
Platform02
Accounting01
07How we delivered
The same three steps as every Codech project: a written proposal, a clickable prototype the client approves screen by screen, then the build. Here are JY Global's real ones.
Highlights from the interactive proposal: the platform, the assistant, the knowledge hub, the client register and the architecture.
Tell us about one workflow you'd like AI to handle. We'll send you a proposal like this one, with the AI design, the screens and the plan, plus an interactive prototype you can click through. Free.
The Managing Director reviewed these 30 pages before the build. New screens are still prototyped first.
Real files from the engagement, with demo data and demo names. Prices have been removed. A few prototype pages (Meetings, Pipeline, Campaigns) are designs for later phases.
Proposal on 14 Jun 2026, first commit on 29 Jun, the portal live on the firm's own server on 6 Jul.
319 pull requests by 30 Sep: forms, two-factor sign-in, IP allow-lists, an audit log viewer and virus scanning.
A CI gate, handover chapters and a test suite of 2,771 cases, so a successor developer can take it on.
08Under the hood
The virus scanner had reported healthy for four weeks but had checked 6 of 377 files, because only the client-form path called it. Scanning moved into the one function every attachment passes through, the size limits are now checked against each other, and the backlog was rescanned.
The client asked whether a role without Reminders could still create one by asking the AI. It could. Now every assistant tool asks for the same permission as its API route, and a role with no grants gets no gated tools at all.
One security header silently blocked PDF previews for four days. The written evidence of seven uncaught failures became a GitHub Actions gate: tests, type checks, lint and build on every change.
Reach out to us now. We'll look at where your deadlines, documents and know-how live today and show you what an AI portal on your own server would change, in a free 45-minute call. Then we send you a free AI proposal and an interactive prototype.
Screens show demo data and demo names from the approved prototype. Prices have been removed from the proposal.