Consensia Agent
Privacy is first
About
The problem it solves # Consensia Agent > Transforming social media noise into verified blockchain intelligence. Consensia Agent is a multi-layered AI system built to solve the **Web3 Information Gap**. In a market where seconds can cost millions, Consensia acts like an autonomous detective, cross-referencing high-speed social signals with immutable on-chain facts. --- ## Core Innovation: Dual-SLM Pipeline Unlike generic LLM wrappers, Consensia uses a cascaded architecture of two specialized Small Language Models (SLMs), each optimized for a specific Web3 task. | Layer | Model | Role | Primary Objective | | --- | --- | --- | --- | | 1 | Tagger (1.5B) | The Sentinel | High-speed stream sanitization | | 2 | Analyzer (3B) | The Strategist | Deep reasoning and investigative tool-calling | --- ## 1) Tagger (1.5B): The Sentinel **Purpose** High-speed stream sanitization. **Function** - Instantly filters out up to 90% of noise: scams, bot spam, and unrelated chatter. - Maps raw text into a strict, structured Web3 taxonomy. **Optimization** - Trained on a custom-built dataset to enforce schema consistency. - Ensures every output is compatible with downstream database queries. --- ## 2) Analyzer (3B): The Strategist **Purpose** Deep reasoning and investigative tool-calling. **Function** - Applies Chain-of-Thought (CoT) reasoning to evaluate Tagger output. - When a signal is high-risk, it autonomously triggers blockchain investigation tools. **Dynamic Context Loop** - `REQUEST_CONTEXT`: fetches missing on-chain data needed for confident decisions. - `DROP_CONTEXT`: discards unconfirmed rumors through negative-learning behavior. --- ## Why It Matters Consensia is designed for high-frequency, high-uncertainty environments where latency and verifiability are critical. The pipeline separates fast filtering from deep analysis, creating a system that is both scalable and reliable for Web3 intelligence workflows.
How it's made
Challenges we ran into # 🛠 Technical Excellence & Challenges ## 🧠 Solving the "Hallucinating Taxonomy" Early iterations of our **1.5B model** invented non-existent tags. We solved this by implementing **Strict Schema Enforcement**. **Approach:** - "Teacher-Student" training approach with GPT-4o - Generated **18,000+** high-fidelity entries - Forced the model to operate within a **fixed, standardized Web3 vocabulary** --- ## ⚖️ Action Disconnect: Reasoning vs. Execution To prevent the model from "guessing" facts before seeing blockchain data, we architected a **Two-Pass Reasoning Loop**: 1. **Decision Pass** — The model determines if a tool call is needed. 2. **Synthesis Pass** — The model generates a final report only after receiving real-time transaction data from the orchestrator. --- ## 🧹 Cognitive Load Reduction (Data Washing) Raw social media data is messy. We implemented a **Pre-processing RegEx Pipeline** that: - Strips artifacts (`&`, bot-tags, broken URLs) - Processes data before it reaches the SLM - **Reduced token consumption** by focusing 100% on economic signals - **Significantly increased classification accuracy** --- ## 📡 Stream-to-Disk Architecture To handle massive data generation without loss: - Moved away from **RAM-heavy processes** - Implemented **Stream-to-Disk** approach - Immediate file flushing after every API interaction - **Ensured data integrity** and **100% resilience** against network failures or rate-limiting --- ## 🌐 Vision: The Future of On-Chain Truth **Consensia** is designed to be **Edge-Ready**: - **1.5B and 3B models** optimized for decentralized compute - Runs on platforms like **SpaceComputer** - The "Truth Engine" of Web3 remains **censorship-resistant** as the blockchain itself