title	emoji	colorFrom	colorTo	sdk	sdk_version	app_file	pinned
ThreadNavigatorAI2.0	🧵	red	green	streamlit	1.33.0	app.py	true

🧵 ThreadNavigatorAI 2.0 — Multi-Agent Reddit Thread Analyzer with LLM-as-a-Judge

🔍 ThreadNavigatorAI 2.0 is a modular, multi-agent Reddit analyzer that summarizes threads, verifies claims, and scores insights using LLM-as-a-Judge. Built for high-scale moderation and discourse comprehension under real-world latency and token constraints.

🎯 Designed for folks evaluating top-tier AI engineering talent.

---

🚀 What’s New in 2.0

ThreadNavigatorAI 2.0 builds directly on v1.0 by introducing:

Feature	1.0	🚀 2.0 Upgrade
Agent Orchestration	LangGraph (Reply, Mod)	Modular pipeline (Summarizer, FC, Eval)
Summarization	Prompt + RAG	Semantic + config-driven transformer
Moderation	Rule-based + LLM	🔍 Fact Checker Agent (tool-augmented)
Evaluation	Manual	🧠 LLM-as-a-Judge (rubric-based)
UI	Basic Streamlit	✨ Tabbed layout, latency toggle, download
Scalability	Single-thread only	100 threads (real + mock hybrid)

---

📊 Architecture Overview

Thread JSON → [Multi-Agent Orchestration]
              ├── 🔍 Summarizer Agent → Extracts core insights
              ├── 🧪 Fact Checker Agent → Verifies claims via tools
              └── 🧠 Evaluator Agent → LLM-as-a-Judge on rubric (Relevance, Factuality...)

              ↳ Config-driven model orchestration (via config.yaml)
              ↳ Latency + model tracking in UI
              ↳ Deployed on Hugging Face Spaces (Streamlit + Docker)

---

🌟 Key Features

✅ Multi-Agent Stack (Summarizer, Fact Checker, Evaluator)
✅ Semantic Summarization + Retrieval
✅ LLM-as-a-Judge via rubric-based evaluation (like RAGAS)
✅ Hybrid Simulation Mode (10 real + 90 mock = 100 total threads)
✅ Live Latency Display + Model Attribution
✅ Streamlit UX with onboarding, source links, toggles, download
✅ Free-tier Deployable (OpenRouter APIs, Streamlit, Hugging Face)

---

🧠 Agent Stack (Config-Driven)

Modular multi-agent architecture modeled after real-world Reddit analysis workflows.

Agent	Role	Model Used (Free-tier)
🧠 Summarizer	Extracts semantic insights from messy Reddit posts using transformer-based compression	moonshotai/kimi-k2:free
🧪 Fact Checker	Verifies factuality using retrieval tools	deepseek/deepseek-r1:free
📊 Evaluator	Scores output on Relevance, Coherence, and Factuality (LLM-as-a-Judge)	openrouter/mistralai/mistral-7b-instruct:free
🔧 Tools	External retrieval & KB (Serper, Wikipedia)	—

All agents use individual OpenRouter models, with latency tracked per call.

---

🔎 How It Works

Select a Reddit-style thread from 100 examples.

Agents are triggered in sequence:

• 🧠 Summarizer → condenses the entire thread
• 🧪 Fact Checker → verifies major claims with external data
• 📊 Evaluator → scores the summary based on LLM rubrics

Output is displayed in a polished tabbed UI with latency + model trace.
Optionally download the result as JSON.

All agents are modular — swap-in / swap-out via config.yaml.

---

🖼️ UI Preview

Tabbed layout includes:

• 🧠 Summary (with model and source)
• 🔎 Fact Checks (claim + judgment)
• ⚡ Latency (per agent)
• 📊 Evaluation (scored by LLM)

🚢 Live Demo

📌 Try the fully working UI on Hugging Face:
👉 ThreadNavigatorAI 2.0 – Live Space

No login, no billing — real inference with OpenRouter free-tier models.

---

🧪 Sample Output (Thread ID: threadid_011)

• 🧠 Summary: Debate around GPT-4 vs Gemini, with user opinions and factual misunderstandings
• 🔎 Claim: “Gemini was trained on YouTube comments” → 🔴 Incorrect

📊 Eval Score:

• Relevance: 🟢 5 — Very relevant
• Coherence: 🟡 3 — Some redundancy
• Factuality: 🟢 4 — Mostly accurate

⏱️ Latency:

• Summarizer: 2.1s
• FactChecker: 3.4s
• Evaluator: 1.9s

---

📁 Project Structure

ThreadNavigatorAI2.0/                      
├── ui/                        
│   └── app.py                     # Streamlit frontend
├── cli/ 
│   └── batch_run.py               # Multi-thread processor
├── config/                        # Models and tool settings
│   └── config.yaml
├── agents/
│   ├── summarizer_agent.py
│   ├── factchecker_agent.py
│   └── evaluator_agent.py
├── utils/
│   ├── llm_utils.py
├── data/
│   ├── batch_output.json          # 10 threads real inference and 90 threads simulated inference
│   └── threads_100.json           # 100 simulated threads
├── assets/
│   ├── threadnavigator_flow.png   # Architecture Diagram
│   └── threadnavigator-demo.gif   # UI Demo
└── requirements.txt

---

💼 Why This Project Stands Out

• ✅ Represents real-world pipeline for Reddit moderation/analysis
• ✅ Dynamic toggles: latency, evaluation, download
• ✅ Hybrid inference simulates scale under free-tier
• ✅ LLM-as-a-Judge with model attribution & traceability
• ✅ Fully local, frictionless UI with onboarding
• ✅ Built like an internal debugging tool for social platforms

---

🧰 Run Locally (for Devs)

git clone https://github.com/rajesh1804/ThreadNavigatorAI2.0
cd ThreadNavigatorAI2.0
pip install -r requirements.txt

Create .env:

OPENROUTER_API_KEY=sk-
SERPER_API_KEY=

Then run:

streamlit run app.py

---

🧠 Linked Projects

Project	Description	Link
🧵 ThreadNavigatorAI 1.0	RAG-based summarizer + moderator + reply agent using LangGraph	🔗 View
🚕 RideCastAI	Real-time ride fare/ETA prediction with spatial heatmaps	🔗 View
🛒 GroceryGPT+	Vector search + reranking grocery assistant with fuzzy recall	🔗 View
🎬 StreamWiseAI	Netflix-style movie recommender with Retention Coach Agent	🔗 View

---

⚠️ Known Challenges

• 🧪 gemma-3-4b-it rejected evaluation due to instruction tuning being disabled — switched to mistralai-7b:free
• 📉 Reddit data volume limited to 10 threads (real) due to OpenRouter limits — mock data generated for scalability tests
• ⏱️ Tool-based fact-checking adds ~2–3 seconds latency — handled with async retries + optional toggle
• 🧵 No full LangGraph orchestration in this version (ThreadNavigator 1.0 had it) — but this version enables more precise per-agent control

---

🧑‍💼 About Me

Rajesh Marudhachalam — AI/ML Engineer with a focus on building LLM-native applications.
📍 GitHub | LinkedIn

Projects: RideCastAI, StreamWiseAI, GroceryGPT+

---

🙌 Acknowledgments

• OpenRouter — Free-tier LLM APIs
• Streamlit — UI framework
• Serper.dev — Web search API
• Wikipedia API — Factual KB
• Hugging Face Spaces — App Hosting

---

📜 License

MIT License

⭐️ Star this repo if it impressed you. Follow for more elite-level ML + LLM builds.