SlackJaw - Your AI-Powered Slack Workspace Assistant

🚀 Inspiration

During hackathons and team collaborations, critical information gets lost in endless Slack threads. Team members ask the same questions repeatedly: "What's the WiFi password?", "What did we work on yesterday?", "Can someone analyze this screenshot?" I built HackBot to solve this - an intelligent Slack bot that remembers everything, understands context, and provides instant answers.

💡 What it does

HackBot is a multimodal AI assistant that lives in your Slack workspace and provides:

🧠 Contextual Memory

Remembers all conversations across channels
Recalls information from days or weeks ago
Provides intelligent answers based on workspace history
Cites sources with direct links to original messages

👁️ Vision Capabilities

Analyzes images, screenshots, and diagrams
Extracts text from photos
Describes visual content in detail
Answers questions about uploaded images

🔗 Document Intelligence

Automatically extracts and summarizes content from shared links
Remembers document contents for future reference
Provides context-aware answers about shared resources

🐙 GitHub Integration

Summarizes recent commits for daily standups
Creates GitHub issues directly from Slack
Tracks repository activity
Keeps team updated on code changes

💬 Smart Conversations

Maintains conversation threads
Understands context from previous messages
Provides relevant, concise responses
Filters out noise and focuses on important information

🛠️ How I built it

Architecture

Backend (Node.js + Express)

Slack Bolt SDK for real-time event handling
Socket Mode for secure, firewall-friendly connections
Multimodal processing pipeline for text, images, and documents

Orchestrator Service

Intent detection and routing
Letta integration for persistent memory
GitHub MCP (Model Context Protocol) integration
Lava Gateway for Routing

AI Stack

Letta: Persistent memory and context management
Lava Gateway: Routing GPT4 for text, gemini-pro for vision, and Claude 3 for code-based queries
GitHub MCP: Direct repository access via Model Context Protocol

Key Technologies

Slack API: Real-time messaging and file handling
Axios: HTTP client for API communication
dotenv: Environment configuration
Model Context Protocol: Standardized tool integration

Data Flow

Slack Message → Backend → Intent Detection → Orchestrator
                                                ↓
                                    ┌───────────┴───────────┐
                                    ↓                       ↓
                            Letta Memory              GitHub MCP
                                    ↓                       ↓
                            Lava AI Processing      Commit Data
                                    ↓                       ↓
                                Response Generation
                                    ↓
                            Slack Thread Reply

🏆 Accomplishments that we're proud of

Multimodal Intelligence: Successfully integrated vision, text, and document processing in a single bot
Persistent Memory: Implemented Letta for true long-term conversation memory across sessions
Smart Citations: Built intelligent source attribution that only cites relevant messages
GitHub MCP Integration: Pioneered use of Model Context Protocol for seamless GitHub access
Performance Optimization: Reduced response time by 60% through intelligent message filtering
Thread Management: Automatic conversation threading keeps discussions organized

📚 What I learned

Model Context Protocol (MCP): Learned to integrate external tools using the emerging MCP standard
Multimodal AI: Discovered the challenges and solutions for processing multiple data types
Memory Management: Implemented efficient context windows to balance memory and performance
Slack API Nuances: Mastered Socket Mode, file handling, and thread management
Intent Detection: Built robust pattern matching for routing queries to appropriate services
Citation Systems: Created smart algorithms to identify and cite relevant sources

🚧 Challenges I ran into

MCP Integration Complexity: GitHub MCP required custom transport layer and careful error handling
Response Time: Initial implementation was slow; optimized by reducing message processing from 100→40 messages
Citation Accuracy: First version cited irrelevant messages; built keyword-based relevance scoring
Image Processing: Handling various image formats and sizes required careful optimization with Sharp
Memory vs Performance: Balancing comprehensive context with fast response times
Orchestrator Routing: Debugging silent failures in the routing layer took significant effort