Real-time sentiment intelligence that speaks fluent crypto.
How we built an autonomous social intelligence platform for MEXC — separating real market signals from bots, slang and coordinated FUD across X, Telegram, Discord and Reddit.
- CLIENT
- MEXC
- INDUSTRY
- Crypto exchange
- MODEL
- Self-hosted Qwen on vLLM
- STACK
- pgvector · Zod · Next.js
Generic social listening can’t read “rekt”.
MEXC needed 24/7 visibility into brand perception, user sentiment and campaign performance across hyper-active, fragmented crypto communities on X, Telegram, Discord and Reddit.
The chatter moves fast and is dominated by slang, bot-driven spam, rapid market swings and coordinated FUD campaigns. Traditional social listening tools couldn’t tell actionable signals from noise — or measure whether a campaign actually landed.
Four layers from raw chatter to a decision.
Streaming ingestion & noise scrubbing
Pipelines across X, Telegram, Discord and Reddit, with deterministic filters for bot spam, duplicate airdrop promos and cashtag hijacking.
Semantic narrative clustering
1024-dim embeddings in pgvector HNSW group chatter into narrative threads in real time — feedback, feature requests and emerging controversies.
Crypto-native sentiment engine
A self-hosted Qwen with Zod-enforced JSON reads HODL, rekt, moon and rugpull to classify Bullish, Bearish, Neutral and Panic/FUD — and attributes spikes to campaigns.
FUD alerts & command center
Influence scoring separates high-impact voices from retail noise; anomaly alerts hit Slack and Telegram, and a live Next.js command center tracks it all.
Two analytics tracks, one private AI core — and alerts within seconds.
- LAYER 1Ingestion routerX · Telegram · Discord · Reddit
- LAYER 2ACampaign analytics#Airdrop · #Listing · KOLs
- LAYER 2BCommunity sentimentpgvector HNSW clustering
- LAYER 3Qwen AI corePrivate vLLM + Zod
- LAYER 4Alerts & command centerSlack · Telegram · Next.js
| LAYER | MODULE | TECHNOLOGY | SCOPE |
|---|---|---|---|
| 1 | Multi-channel ingestion & filter router | API streams & webhooks | X, Telegram, Discord, Reddit; bot spam pruning & cashtag dedup |
| 2A | Campaign & post analytics | Marketing analytics pipeline | Campaign hashtags, KOL amplification & virality metrics |
| 2B | Brand & community sentiment | Real-time NLP + pgvector HNSW | Mood index, FUD anomaly detection, narrative clustering |
| 3 | Self-hosted Qwen AI core | Private vLLM + Zod schemas | Crypto-slang understanding, campaign attribution, structured JSON |
| 4A | Real-time operational actions | Automated dispatch | Instant Telegram & Slack alerts on FUD spikes or sentiment shifts |
| 4B | Executive command center | Next.js real-time analytics | Live sentiment, campaign ROI, competitive share of voice, heatmaps |