Ask Hn: My Multi-agent Financial Sentiment Architecture

Hello HN,I’m a college student building a financial intelligence terminal for personal use. The goal is to capture "whisper numbers" and sentiment shifts in social media before they hit mainstream headlines.I would love your feedback on the architecture and logic.The Stack:Backend: Python (PM2 managed)LLM: OpenAI GPT 5.2 for high level analysis + GPT-5-mini for news filter and market data collection (via API)Frontend: Next.js + Tailwind + RechartsThe Architecture (The interesting part):1. The CollectorInstead of hitting the LLM immediately, I run a raw scraper every 30 minutes targeting:- Nitter instances (for Twitter/X data without API limits).- GNews RSS (for official headlines).- DuckDuckGo (for general forum chatter).- 1-min OHLC Data (via yfinance) to monitor price micro-structure.2. The Agentic SystemI split the analysis into two roles:- Market Agent: Analyzes the 30-minute window of 1-min candle data. It looks for patterns like "V-shape recovery," "Flash crash," or "Volume exhaustion" (things a simple % change metric misses).- News Agent: Analyzes the messages and events from the scrapper, and it can determine whether to wake up on-call senior agent for "emergency" during market open time.- Senior Agent: Receives the cleaned news stream + the Market Agent's technical summary. It produces a Sentiment Score (-10 to +10) and a rationale. It creates a "red alert" if the price will probably be changing a lot in the next several hours.- On-call Agent: Analyzes the news and search the internet for proofs and determine whether to send alert to users or not.3. The FrontendTo better display information, I vibe coded a Next.js + Tailwindcss frontend. An updated version with i18n will soon be published.My Questions for HN:Latency vs. Depth: Currently, the AI analysis cycle takes ~3 minutes. For "swing trading" this is fine, but will it help if there's a better pattern to stream partial updates to the frontend without waiting for the full analysis to complete?Hallucination Risks: I force the LLM to verify dates, but sometimes it still treats a "re-posted old news" as a new event. How do you guys architect verification layers for news agents?Practicability: I just finished this agentic system, and it's my first day trading according to the info from the agentic system and it's currently +0.8% profit. I'd love advice from someone with more trading experience. Also, feel free to reach out to me if you want a preview of the system's output.Some screenshots here: https://x.com/CLCKKKKK/status/2006085046269337799?s=20My email: yiz29@illinois.eduThanks!