- Domov
- Komunita
- Programovanie
- Built an AI trading platform for my intraday trading, what I...
Built an AI trading platform for my intraday trading, what I've learned
So I've been building this for my own day-to-day trading for a while now, and I figured it's time to share. Not trying to sell anything — I just couldn't find a tool that did what I actually wanted, so I built it myself. Would love feedback from other traders who've faced the same frustrations.
The idea started simple: I wanted an AI that analyzes charts the way I think about them as a trader, not just pattern-matching on historical data. Ended up being way more complex than expected.
The core thing I built is a two-agent pipeline — one agent reads market structure (supply/demand zones, key levels, overall bias), the other takes that context and looks for actual trade setups with entry, SL, TP and R:R. What I liked about this approach is that the analysis is transparent — you can literally watch the reasoning unfold in real-time, and the zones get drawn on the chart as the agent identifies them. It uses Claude under the hood, mostly because the chain-of-thought quality was noticeably better for this kind of structured reasoning.
The chart side is built on Lightweight Charts v5 — custom drawing tools, cross-timeframe persistence, drag-to-move SL/TP, right-click to place orders at any level. That part took forever to get right honestly.
I also added a paper trading mode that mirrors live trading exactly (server-side tick engine, auto-close on SL/TP) — mostly because I wanted to test my own strategies without blowing real money before going live.
For live trading, I built my own MT5 bridge (desktop companion), which connects to the backend platfom directly. I fetch live and historical OHLC from MT5 that are rendered real-time in the UI, and can directly open/modify/close all my trades directly from the main UI, without having to check on MT5 anymore.
I also built a conversational assistant that runs alongside the analysis. The idea was to have something I could actually talk to about what's happening on the chart — not just run a one-shot analysis and stare at the output. It has full context: my open positions, the latest analysis results, my trading profile. I can ask things like "is this still a valid setup given the last candle?" or "what's my current exposure on gold?" and get a relevant answer. It can also fetch live prices and pull up indicator data on demand. Honestly this ended up being more useful day-to-day than I expected — having that back-and-forth while a trade is running changes how I think about position management.
The personalization layer is the part I'm most unsure about. The idea was to use RAG to make the AI actually aware of my trading context — not just generic market analysis. During onboarding, the system captures my trading style, risk profile, timeframe preferences, and favorite assets. That data gets embedded using Voyage AI's voyage-finance-2 model (finance-specific embeddings, noticeably better than generic ones for this use case) and stored in pgvector. When the AI runs an analysis, it retrieves the most relevant chunks of my profile and injects them into the prompt — so it knows I trade XAUUSD on intraday timeframes with a specific risk per trade, rather than giving me a one-size-fits-all output. It also handles lot sizing automatically based on my account balance and risk %. In theory it's exactly what I wanted. In practice I'm still calibrating how much it actually shifts the analysis versus a well-crafted static prompt. Would be curious if anyone has gone down this road.
Just shipped basic price alerts. The next thing I'm working on is having an advanced alerting system able to detect when price enters an identified zone and there's a confirming candlestick pattern — so instead of a raw price alert I get an actual confluence signal before I even look at the chart.
Tech Stack if anyone's curious:
Next.js + Flask + PostgreSQLpgvector for RAG storage, Voyage AI (voyage-finance-2) for embeddingsClaude for the AI layerTwelve Data for the feed, MT5 bridge for live execution
I'm pretty proud of what I've built — took a lot of trial and error to get here. Now I'm curious to discuss with other traders: does this kind of AI-assisted workflow resonate with you? And if you've tried building or using something similar, what worked, what didn't?Below some screenshots of the UI:




