VerasRecon
Decode the real sentiment behind investor chat groups using a multi-agent LLM pipeline.
Analysis Modules
Main Interface
Central command for real-time operation monitoring.
The Verdict
Final execution decision based on multi-factor synthesis.
Emotional State
Real-time market sentiment quantification.
Cultural Biases
Detection of region-specific behavioral patterns.
Final Reasoning
Transparent chain-of-thought explanation.
3-Phase LLM Pipeline Architecture
The Problem
Reading thousands of messages in trading groups is time-consuming (10k+ messages/day), noisy (80% is memes and off-topic), and misleading (what traders SAY often contradicts what they DO). Traditional sentiment tools miss sarcasm, regional slang, and the critical gap between verbal extremism and real actions.
Technical Challenges
Sarcasm & Irony
"Sure it'll drop 🙄" must be correctly interpreted as frustrated/bullish, not bearish.
Regional Calibration
Argentine financial slang (MEP, CCL, carry trade, "al horno") requires custom lexicon.
Action vs Words
Detecting when verbal panic is high but actual sell actions are zero.
My Solution
- CIS v5 parser: Phone anonymization, date normalization, 60% noise reduction
- Phase 1 (Emotion): Dominant emotion, sarcasm detection, toxicity scoring
- Phase 2 (Biases): Dollar-centrism, crisis memory, carry trade anxiety detection
- Phase 3 (Synthesis): ACTIONS > INTENTIONS > WORDS rule, contrarian signal calculation
- Multi-model fallback: Groq → OpenRouter → Google AI (5 LLMs per phase)