AI Architecture (Tech Overview)

Sentra’s intelligence layer is built on a robust multi-stage AI pipeline designed specifically for understanding crypto-native sentiment in real time. Unlike general-purpose NLP systems, Sentra’s architecture is optimized to process emotional signals from short, slang-heavy, often speculative content found across crypto platforms.


1. Data Ingestion Layer

Sentra aggregates massive volumes of data from both structured and unstructured sources:

  • Social platforms: Twitter/X, Discord, Telegram, Reddit

  • News outlets: Aggregated RSS feeds, crypto media, headline sentiment

  • On-chain data: Smart contract comments, token transfers with social annotations

  • Community behavior: Frequency of keyword repetition, emoji usage, sentiment clusters

This layer operates with high frequency, continuously updating input streams with minimal latency.


2. Natural Language Processing (NLP) Engine

At the heart of Sentra is a custom-trained NLP model using transformer-based architectures (BERT, RoBERTa variants) fine-tuned on crypto-specific corpora.

  • Detects emotional tone (fear, greed, doubt, hype, sarcasm, etc.)

  • Parses slang, memes, abbreviations, and emojis with contextual meaning

  • Scores each message based on polarity, intensity, and emotional certainty

  • Can process multilingual content, prioritizing English, Korean, and simplified Chinese


3. Token Mapping & Signal Routing

After sentiment scoring, messages are contextually matched to tokens, projects, sectors, or influencers:

  • Uses hashtag and keyword association + knowledge graph (e.g. $SOL ↔ Solana ↔ Ecosystem tags)

  • Filters spam and unrelated noise using attention-weighted filters

  • Routes relevant sentiment signals into token-specific scoring pipelines


4. Sentiment Scoring Engine

Each token receives a Sentiment Score (0–100) based on weighted aggregation of:

  • Emotion type (fear/greed/hype)

  • Source trust level (verified vs unverified accounts)

  • Message frequency and clustering behavior

  • Historical correlation with price movements (machine-learned)

This engine constantly learns and adapts using reinforcement feedback from actual market outcomes.


5. Alert Triggering & Visualization Layer

Once emotional thresholds are crossed, Sentra generates alerts and renders visual insights:

  • Alert triggers: Sudden shifts in score, rumor spikes, divergent behavior

  • Visual output: Heatmaps, emotion trendlines, sentiment comparison dashboards

  • Delivery system: Telegram Mini-App UI, web dashboard, notification API


Architecture Layer

Function

Data Ingestion

Collects multi-source social and on-chain sentiment data

NLP Engine

Interprets tone, slang, memes, and emotion from text and emojis

Token Signal Routing

Maps content to relevant tokens, filters irrelevant noise

Sentiment Scoring Engine

Assigns real-time scores using multi-factor emotional modeling

Alert & Visualization System

Generates alerts and renders sentiment data through UI components

Last updated