title	emoji	colorFrom	colorTo	sdk	sdk_version	app_file	pinned
ThreadNavigatorAI	🧵	gray	indigo	streamlit	1.33.0	app.py	true

🧵 ThreadNavigatorAI — Reddit-style Summarizer, Moderator & Reply Assistant (Agentic RAG)

🔍 ThreadNavigatorAI is an intelligent Reddit-style assistant that summarizes threads, detects moderation issues, and suggests replies — powered by LangGraph-based Agent Orchestration + RAG.

🎯 Built for Reddit, Discord, and social platforms that need smart thread navigation at scale.
🛠️ Uses only free-tier tools (OpenRouter, Weaviate, Streamlit, Hugging Face).

---

🧠 What It Solves

"This thread is messy — what’s the summary, what’s toxic, and how should I respond?"

Online communities face:

• Too many replies, not enough signal
• Toxic or biased comments
• New users unsure how to engage

ThreadNavigatorAI solves this by:

• 📝 Summarizing the core thread
• 🛡️ Detecting moderation concerns (bias, sarcasm, trolling)
• 💬 Generating helpful replies

---

🔧 Key Features

✅ Multi-agent workflow (Summarizer, Moderator, Reply Assistant)
✅ LangGraph orchestration with stateful memory and retry logic
✅ Vector-based RAG using Weaviate + Sentence-BERT
✅ Manual Evaluation Framework for recruiters + QA
✅ Live LLM Latency Tracking in UI
✅ Streamlit UI with branding, sample queries, and visual appeal
✅ Free-tier deployable on Hugging Face

---

🧠 Agent Workflow Breakdown

Each user query activates a 3-stage Agentic RAG pipeline orchestrated via LangGraph.

Agent	Role
📝 Summarizer	Uses RAG to summarize key takeaways across the entire thread
🛡️ Moderator	Flags sarcasm, trolling, biased or off-topic replies
💬 Reply Assistant	Crafts helpful, polite, and context-aware replies for the user

All agents share memory via LangGraph's StateGraph — allowing responses to influence downstream decisions.

---

🖼️ Architecture Overview

User Query → [LangGraph Agentic Workflow]
              ├──> Summarizer Agent (RAG over thread)
              ├──> Moderator Agent (LLM + heuristics)
              └──> Reply Assistant Agent (LLM response suggestion)

              ↳ Shared Memory + Retry Handling via LangGraph

All backed by:
- Weaviate Vector DB (cloud-hosted)
- Sentence-BERT embeddings
- Mistral-7B (OpenRouter, free-tier)

---

🧪 Example Flow

A walkthrough for one simulated thread

🧵 Thread 002 – Gemini vs ChatGPT

Query:

"Summarize this thread and point out any sarcasm or trolling."

Agent Flow Output:

• 📝 Summary:
  Debate between Google’s Gemini and OpenAI’s ChatGPT. Highlights bias and opinions.
• 🛡️ Moderation Report:
  Detects sarcastic "toaster" comment and subtle trolling in LLM vs AGI debate.
• 💬 Suggested Reply:
  "Both models have strengths. Let’s stay focused on the facts instead of throwing shade."

⏱️ Latency: 3.2 seconds (displayed in UI)

---

🧭 Manual Evaluation Samples

Manually curated QA pairs for demo/testing/evaluation. Integrated directly into the Streamlit UI.

Thread ID	Scenario	Agent Task
thread_001	Apple M4 chip launch	Sentiment + Summary
thread_002	Gemini vs ChatGPT	Detect sarcasm/trolling
thread_003	Student loan forgiveness	Highlight bias/emotion
thread_004	Apple Vision Pro	Off-topic derailment detection

Used in UI to demo RAG + LLM + moderation workflows under realistic edge cases.

---

🧪 Live Demo

🎯 Try the full agentic flow now on Hugging Face:
👉 ThreadNavigatorAI Demo

---

📁 Project Structure

ThreadNavigatorAI/
├── app.py                         # Streamlit UI
├── eval_samples.py                # Manual eval dataset
├── scripts/
│   ├── thread_uploader.py         # Uploads simulated Reddit threads to Weaviate
│   ├── summarizer_agent.py        # Summarizer agent logic
│   ├── moderator_agent.py         # Moderator agent logic
│   ├── reply_agent.py             # Reply generator logic
│   └── agent_graph.py             # LangGraph agent orchestration
├── assets/
│   ├── threadnavigatorai-architecture.png
│   └── threadnavigatorai-demo.gif
├── requirements.txt
└── README.md

---

💼 Why This Project Stands Out

✅ Agentic RAG + LangGraph orchestration
✅ Evaluation-ready UI with real data
✅ Built like an internal tool for Reddit or Discord moderators
✅ No OpenAI costs — uses OpenRouter free-tier only
✅ Latency, UX, eval — all production-quality
✅ Demonstrates end-to-end agent reasoning + vector retrieval

🎯 Built to impress top-tier AI/ML recruiters, not just LLM hobbyists.

---

🧱 CHeckout My Other Projects

Project	Description	Skills Showcased	GitHub Repo
🛒 GroceryGPT+	AI-powered grocery search engine with vector DB, reranking, and typo-tolerant recall	Semantic search, LLM reranking, Weaviate, Sentence-BERT, OpenRouter	🔗 Repo
🚕 RideCastAI	Predicts ride fare & ETA with dynamic spatial heatmaps and real-time latency visualization	Spatio-temporal ML, regression, simulated demand mapping, latency-aware UI	🔗 Repo
🎬 StreamWiseAI	Netflix-style movie recommender with a Retention Coach agent and session-aware personalization	Recommender systems, RAG agent, session memory, fuzzy search, real-time LLM retry handling	🔗 Repo

---

🚀 Getting Started

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

Create a .env file with:

WEAVIATE_URL=your_cluster_url
WEAVIATE_API_KEY=your_api_key
OPENROUTER_API_KEY=your_openrouter_key

Then:

streamlit run app.py

---

💡 Skills Demonstrated

✅ LangGraph agent orchestration
✅ Modular agent design (summarization, moderation, replies)
✅ Free-tier LLM usage (OpenRouter)
✅ Weaviate cloud vector DB + RAG over Reddit-style threads
✅ Manual eval integration in UI
✅ Latency tracking and real-time feedback
✅ Streamlit UI with product thinking and branding
✅ Fully deployable on Hugging Face Spaces

---

🧑‍💼 About Me

Built by Rajesh Marudhachalam

• AI/ML Engineer UofT CS
• GitHub: github.com/rajesh1804
• LinkedIn: linkedin.com/in/rajesh1804

---

🙌 Acknowledgments

• LangGraph for stateful agent workflows
• OpenRouter for free-tier LLM APIs
• Weaviate Cloud for vector storage
• Hugging Face Spaces for frictionless deployment
• Sentence-BERT for embeddings

---

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