Ask HN: My multi-agent financial sentiment architecture
2 comments
Hey, for less latency for the llm portion, take a look at Cerebras. It won't run OpenAI models, but if you can substitute for an equivalent you might be able to have better speeds. There is a memory constraint so not sure if it's the most suitable for the project though. Curious to see how this works out in terms of consistent durable performance.
Ok I will. Thanks for your suggestion. It doesn't matter if it's from OpenAI.
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 + Recharts
The Architecture (The interesting part):
1. The Collector
Instead 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 System
I 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 Frontend
To 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=20
My email: [email protected]
Thanks!