Elfbot — Human-in-the-Loop AI for Comment Moderation & Disinformation Triage
From LLM-agent disinformation triage (v1) to AI-based Facebook comment moderation (v2)
Elfbot is a human-in-the-loop AI system developed at logiq.media, where I am the lead developer. The current version (v2) focuses on AI-based Facebook comment moderation: a backend for ingestion, polling, and webhooks, and a core pipeline combining classifier rules with fine-tuned LLMs to detect trolls and toxic comments — with human review kept in the loop rather than auto-flagging (repo private at the moment).
The first version (v1, linked below) was broader in scope, built for digital-trace analysis and disinformation triage.
This first version uses LLM agents to query multiple evidence sources — Wikidata, Tavily, and DuckDuckGo — to assemble multi-source evidence and produce citation-linked, confidence-calibrated outputs at the level of individual claims. Coordinated-behaviour signals such as burst dynamics and account clusters (on Bluesky and X/CSV samples) are surfaced for human review rather than auto-flagged.
Stack: Python, LLM agents, MCP, retrieval (Tavily/Wikidata), reproducible pipelines.
Code: github.com/nimathing2052/elfbot_prototype. demo. Pitch Video(presented by my Colleague, Tom).