AI Modules

PolinaOS runs on a multi-agent AI system. Each agent is specialized and communicates through structured JSON outputs and verifiable pipelines. Together they form the closed-loop intelligence layer that powers task automation and reward allocation.

---

1. 🧠 Analyst Agent – Polina

AI researcher & context builder

• 📡 Context Loading → Scrapes Twitter timelines, contract mentions, on-chain events. → Uses Puppeteer + RPC clients to rebuild the project’s narrative landscape.

• 🗂️ Knowledge Memory → Pulls past missions + contributor data from PolinaOS DB. → Enables incremental analysis (no re-learning from scratch).

• 📊 Signal Collection → Finds top tweets, memes, influencer replies, engagement baselines. → Detects abnormal attention spikes via statistical thresholds.

• 🧾 Project Understanding Generation → Outputs structured JSON:

  {
    "strengths": [...],
    "weaknesses": [...],
    "activation_opportunities": [...]
  }

  → Feeds into Creator Agent.

---

2. ✨ Creator Agent – Lara

AI task designer, optimized for Twitter virality

• 🎯 Mission Framing → Selects high-value tweets/narratives for amplification. → Uses contextual embeddings to detect “shill-worthy” content.

• 📝 Task Design → Generates prompts with hashtags, media, reward rules. → Embeds cultural priors (CT memes, slang, timing windows).

• 📦 Dynamic Packaging → Outputs tasks in schema:

  {
    "objective": "...",
    "deadline": "...",
    "kpis": ["views", "engagements"],
    "tiers": ["basic", "advanced", "viral"]
  }

• 🔄 Adaptive Iteration → Monitors early submissions. → Auto-generates A/B variants of prompts for higher engagement.

---

3. 🎯 Scoring Agent – Zero

Autonomous evaluator & reputation engine

• ✅ Task Matching → Uses semantic search + regex/rule validation. → Verifies tweet/post contains required hashtags, media, timing.

• 🧮 Effort & Originality Scoring → LLM-powered analysis of tone, length, novelty. → Penalizes template spam & rewards authentic, thoughtful content.

• 🚫 Sybil & Spam Filtering → Cross-checks:

  • Wallet activity (on-chain)
  • Twitter history
  • Submission duplication → Blacklists repeated farming patterns.

• 🧠 Reputation Memory → Maintains contributor history:

  • score_total
  • score_avg
  • sybil_flags → Adjusts reward weighting dynamically.

• 📤 Output Schema

  {
    "contributor": "0x...",
    "task_id": "12345",
    "score": 87,
    "flags": ["sybil_suspected"],
    "notes": "Low effort retweet"
  }