An Arabic-first AI trading analyst, live on Google Play with paying subscribers.
Overview.
Arabic-speaking retail traders
Gulf-dialect users trading gold, forex, crypto and US indices on MT4, MT5 and TradingView. Most existing tools are English-only, expensive, and assume card payments that many of these users do not have.
Analysis takes an hour and still feels like guessing
Reading structure, marking order blocks, sizing the position and keeping discipline is slow, error-prone and emotional. Signal groups on Telegram are opaque and unverifiable.
One tap to a validated trade plan
A native app where the AI does the technical read, the math is checked before it reaches the user, position sizing is built in, and a voice analyst answers follow-ups. Sold through tiers and coins, paid by bank transfer with in-app approval.
Feature walkthrough.
A home screen that reads the market for you
The first screen answers three questions in a glance: what session is open, how the market feels, and what the assistant thinks you should look at right now.
- 1Session awarenessDetects the active trading session (Asian, European, US) and warns when an index market is closed so the analysis will rely on yesterday's close.
- 2Mascot nudgeThe Spirex mascot surfaces a one-line brief, gold's move today for example, and offers to open the full analysis or ask the AI assistant with one tap.
- 3Market mood and moversA sentiment gauge and a counter of strong 24-hour moves, computed from live prices without spending an analysis credit.
- 4Voice assistant entryA live card that starts a real-time voice call with the AI analyst from the home screen.
- 5Below the foldThe Spirex Vision card, a Gold Now card with live XAU/USD price and momentum, a 24-hour chart with volume bars, market chips for EUR/USD, BTC and ETH, and the daily discipline tip.
From a pair and a timeframe to a complete trade plan
Pick gold, silver, a forex pair, a crypto pair or a US index. The setup screen shows what the engine already knows, you choose the timeframe and the discipline style, and the backend fetches candles, computes the indicators server-side and asks GPT-4o for a structured signal.
- 1Engine status and market readBefore spending a credit the screen shows whether the gold engine is online, plus live liquidity, volatility and chart-trend reads.
- 2Timeframe pickerDaily, 4 hours, 1 hour or 15 minutes. The analysis is multi-timeframe underneath, so the pick sets the horizon, not the context.
- 3Analysis styleStandard says WAIT when there is no clear setup. Take the risk pushes for a direction. Each style has its own discipline rules server-side.
- 4What the engine usesSeven technical indicators, ten years of history and four timeframes, all computed on the edge function so the model receives facts, not raw candles.
- 5The resultBUY, SELL or WAIT with a confidence score, entry zone, stop loss, two targets and risk-reward, validated by a geometry guard that repairs or downgrades broken setups, then saved to My Analyses.
Talk to the analyst, live, in Gulf Arabic
A speech-to-speech call over WebRTC to OpenAI's realtime model. The assistant has a personality, knows the markets, and can look things up mid-sentence.
- 1Real-time voice, interruptibleThe status pill shows the session state. Audio is low-latency WebRTC with Arabic transcription, and you can cut in while it is speaking.
- 2Live tools during the callFunction calls are relayed to a Supabase Edge Function that fetches live prices, FX rates, news, the Fear & Greed index and top movers, then feeds the answer back into the session.
- 3Text chat with tier routingThe keyboard button switches to the same persona in text, routed to GPT-4o or GPT-4o-mini by subscription tier with daily message limits.
- 4Metered and abuse-resistantThe entitlement pill reflects the server's decision. Trial minutes are bound to device id plus email, sessions carry a server-issued second budget, and 30-second heartbeats stop force-quit from dodging the meter.
Screenshot any chart, get a setup back
Traders already live in MT4, MT5 and TradingView. Vision meets them there: upload a chart screenshot and the model reads it.
- 1Three steps, no broker integrationPick, Analyze, Result. A screenshot from MT4, MT5, TradingView or any app is enough.
- 2Reads pair, timeframe and priceOpen photos or take a picture with the camera. GPT-4o vision extracts the instrument and levels before analysing the structure.
- 3Returns entry, stop and targetsThe output uses the same signal format as the standard analysis, so the lot calculator works on it too. A guide explains how to take a screenshot the model can read.
Capital protection built into the flow
Most retail losses come from position size, not direction. Spirex puts the risk math one tap away from every signal.
- 1Risk GuardEnter balance, risk percentage (1 to 3 percent) and leverage (1:100 to 1:1000). The app recalculates exposure live.
- 2Calculate my lot and executeAfter an analysis, one button sizes the contract from your balance and risk, using the signal's stop distance.
- 3Discipline tipsShort daily reminders on the home screen about emotion and rules.
Every market gets its own engine
The Signals tab is the hub: pick a market, see what your tier unlocks, and start. Around it sit the community layer, a live win ticker of results and a Telegram channel.
- 1Tier and live statusThe badge shows the active plan and whether live market data is flowing.
- 2Featured marketsGold is the flagship engine with unlimited analysis on PRO and above. Silver, forex, indices and crypto each have their own card and gate.
- 3Crypto VIP strategiesExclusive RSI, MACD and EMA strategies on Binance data with a 30-minute cooldown per pair to keep analysis quality high.
A dual-coin economy and manual payments that scale
Spirex sells in markets where card payments are unreliable, so the monetisation had to work without a payment gateway. The profile is the wallet.
- 1Plan badge and legacy-safe tiersSixteen historical subscription ids map to today's buckets, so a paying user from an old version never loses access.
- 2Gold balanceGold coins unlock gold analysis and VIP strategies. Unlimited from PRO up, so the balance reads infinity.
- 3Silver pointsSilver coins pay for regular analyses. New users get a welcome bonus, and a coin is refunded automatically when an analysis fails technically.
- 4Subscription details and paymentsActivation date, coins used and plan management. Upgrades go through a payment-proof flow: upload the transfer receipt, watch a countdown for the review window, and an admin approves it. Entitlements are written server-side only.
Operations without a laptop
An in-app admin role manages the business from the same build users run. The user list is blurred here for privacy.
- 1Search any userBy name, email or phone, then change tier, top up coins or set a role in place.
- 2Live KPIsSubscribers per tier, free users, total accounts and total revenue, straight from Postgres.
- 3Approvals and the kill switchReview uploaded payment proofs, and use the app_updates table to push every installed client to the Play Store when an endpoint changes.
Architecture.
Engineering deep dives.
1. Porting a visual n8n pipeline to serverless, and hiding the keys
- Problem
- The analysis backend ran on a self-hosted n8n VPS with API keys inside workflow JSON, reachable over plain HTTP.
- Approach
- Rewrote it as a Deno edge function: an asset-class router, a provider chain with fallbacks (Twelve Data, then Binance PAXG proxy or Yahoo, silver via SI=F futures), and an indicator engine (EMA, RSI, MACD, ATR, fractal swings, FVG detection) that runs server-side so the model receives facts, not raw candles. Keys moved to Supabase Vault behind service-role-only SECURITY DEFINER RPCs.
- Trade-off
- Kept the old HTTP-200-with-success:false error contract so already-shipped clients kept working.
2. A geometry guard for numbers people trade on
- Problem
- The model occasionally inverted stop and target levels or drifted far from the live price. Unacceptable when users act on it.
- Approach
- Post-generation invariant enforcement: a sanity band of ±25% around live price, directional ordering rules per BUY/SELL, ATR-based reconstruction of broken levels, recomputed risk-reward with a mode-aware floor that downgrades the signal to WAIT, and a confidence cap whenever a repair happened.
- Trade-off
- A repaired setup can differ from the model's own reasoning text.
3. Voice-call cost and abuse control
- Problem
- Realtime voice costs about $0.126 per minute, and a free trial is trivially farmable with new accounts.
- Approach
- A one-time trial ledger keyed on the union of device id and email, server-side entitlement checks that mint short-lived OpenAI Realtime client secrets, a hard per-session second budget returned to the client, 30-second usage heartbeats, and consumption decided from real subscription state, never from what the client claims.
- Trade-off
- iOS vendor id rotates after a full uninstall. Accepted, and compensated by the email leg.
4. WebRTC speech-to-speech with server-side tools
- Problem
- OpenAI Realtime sessions cannot call the app's own data tools by themselves.
- Approach
- A React Native WebRTC peer connection to gpt-realtime. Function-call events are relayed by the client to the edge function as run_tool requests, which execute live-data fetches and return results into the session. Android audio routing quirks after route changes are handled with an in-call manager.
- Trade-off
- The client-relayed tool loop adds one round trip per call.
5. Tier migration without breaking paying users
- Problem
- Sixteen subscription ids accreted across versions. A naive bucket check once gave 58 paying PRO users a zero cap.
- Approach
- A single mapping layer (bucketForSubscription, isProTier, isVipTier, getHourlyLimit) with explicit legacy-id lists kept in sync between app and edge function. Per-tier hourly analysis limits, daily chat limits and per-message character caps applied to every message in history, which also blocks spoofed-history cost attacks.
- Trade-off
- Tier lists are duplicated client and server and must be kept in sync by hand.
6. Force update and remote kill switch
- Problem
- Shipped clients pointing at retired endpoints need to be movable without waiting for organic updates.
- Approach
- An app_updates table is checked at launch and drives a blocking update screen to the Play Store link. Each release commits its own SQL toggle to arm or disarm the gate.
- Trade-off
- Manual SQL toggles, no staged rollout.
Business model.
- Welcome-bonus silver coins
- Free crypto analyser (50+ coins)
- Free gold trial
- Text chat on GPT-4o-mini
- More silver coins
- Wider analysis access
- Higher hourly limits
- Unlimited gold coins
- Gold and silver VIP analysis
- Crypto VIP strategies, indices
- GPT-4o chat
- Everything unlocked, lifetime
- Spirex AI voice calls
- Uncapped usage
Two currencies, one wallet
Silver coins are spent on regular analyses and refunded when a run fails. Gold coins gate premium features and are unlimited from PRO up. Rate limits are enforced per tier on both client and server.
Payments without a gateway
Users pay by bank or mobile transfer, upload the receipt in the app, and see a countdown for the review window. An admin approves from the dashboard and the entitlement is written server-side. Nothing in the client can grant itself a tier.
Results.
Live, rated 4.8, paying users
Published on Google Play with 56 releases through v8.3.0, more than 28,900 registered accounts in the admin dashboard, a Telegram community, and a paying subscriber base across the Starter, PRO and Infinity tiers.
Server-side metering on every AI feature
Hourly analysis limits, daily chat limits, per-message caps, voice second budgets and device-bound trials keep OpenAI spend predictable and make free-tier farming uneconomic.
Zero servers to babysit
After the n8n port there is no VPS. Two edge functions, a Postgres and Vault handle everything, and a force-update gate can move every installed client in minutes.
What's next.
Engineering debt I would pay first
Crash reporting and analytics (none today), a CI pipeline for signed builds, real test coverage beyond the boilerplate Jest test, and moving the Vision analysis behind the same edge-function and Vault pattern as everything else.
Product roadmap
App Store release from the already-configured iOS project, forex analyser out of coming-soon, staged rollouts for the update gate, and a proper payment gateway alongside the receipt flow for markets that support it.