Work
Maurice, first generation (2024 to 2025): the original web sales assistant
This describes the first version of Maurice. The current Maurice runs on AgentOS and Black Cloud; see the document "About Maurice" for what he is today. This history matters because it is what Blacksky learned from before building the platform.
Overview
Maurice is Blacksky's production AI sales assistant - a conversational AI that qualifies leads, answers questions about Blacksky's services, and captures prospect information through natural dialogue. He's live at blackskymedia.org, handling real sales conversations 24/7. Maurice demonstrates that Blacksky doesn't just talk about AI - we build and deploy production AI systems.
What Maurice Does
Maurice handles the entire top-of-funnel sales conversation:
- Answers questions about Blacksky's services, projects, and capabilities
- Qualifies visitors based on intent signals in their messages
- Scores leads as hot, warm, or cool based on buying signals
- Captures contact information through natural conversation flow
- Remembers returning visitors and picks up where conversations left off
- Escalates qualified leads to the sales team
Maurice's Personality
The first Maurice had a deliberately quirky personality (running jokes, quotes, musical tastes). The current Maurice keeps the dry wit and the self-awareness and dropped the bits: a public phone line for prospects needs a sharp, warm voice more than a routine.
Technical Architecture
Two-Repo Strategy - Same codebase runs locally and in production with different infrastructure:
| Environment | LLM | Vector DB | Database | Hosting | |-------------|-----|-----------|----------|---------| | Production | Together AI (Llama 3.1 70B) | Pinecone | PostgreSQL | Railway | | Local Dev | llama-cpp-python (Llama 3.1 8B) | ChromaDB | SQLite | localhost |
Same prompts, same personality, different infrastructure. This lets Blacksky develop locally on a laptop and deploy to cloud without code changes.
RAG Pipeline
- User asks a question
- Embed query using Sentence Transformers (all-MiniLM-L6-v2)
- Query vector database for relevant Blacksky knowledge (projects, services, case studies)
- Build prompt with retrieved context + Maurice's personality instructions
- Call LLM (Llama 3.1 70B via Together AI in production)
- Stream response back to user in real-time
- Score lead based on intent signals detected in conversation
- Store conversation, lead score, and contact info in PostgreSQL
Key Features
- Streaming responses - Real-time typing effect, not waiting for full response
- RAG retrieval - Accurate information about Blacksky pulled from knowledge base, never hallucinates
- Lead scoring - Automatically classifies visitors as hot, warm, or cool based on buying signals
- User memory - Recognizes returning visitors and continues previous conversations
- Semantic fact extraction - Automatically extracts role, budget, timeline, industry from conversations
- Admin dashboard - View all conversations, lead scores, contact info, and analytics
- Voice AI phone calls - Maurice answers real phone calls with streaming TTS, ambient audio, and persistent caller identity (see maurice-voip-phone-ai for details)
- Phone number identity - Callers are recognized by their phone number across all interactions — like a cookie that never clears
- SaaS platform - A multi-tenant version where customers deployed their own AI assistant with custom knowledge bases (since retired)
Technologies
| Component | Technology | |-----------|------------| | Backend | Python, FastAPI | | LLM | Llama 3.1 70B (Together AI) | | Embeddings | Sentence Transformers (all-MiniLM-L6-v2) | | Vector Database | Pinecone (production), ChromaDB (local) | | Database | PostgreSQL (Railway), SQLite (local) | | Frontend | Vanilla JavaScript | | Hosting | Railway |
Sample Conversation
Visitor: "Why Blacksky, why not one of the big boys?"
Maurice: "We're not a big shop fitting a square peg into a round hole. We're a boutique firm that builds the thing and ships it."
Why This Project Matters
Maurice is proof that Blacksky builds production AI, not just prototypes. He handles real sales conversations around the clock, qualifies leads, and demonstrates the full stack:
- LLM Integration - Production deployment with Together AI
- RAG Architecture - Knowledge retrieval that keeps responses accurate
- Lead Qualification - AI that actually drives business outcomes
- Dual Environment - Same code runs locally and in cloud
- Real Users - Not a demo, a working system serving actual visitors
For prospects, Maurice answers the question: "Can Blacksky actually build AI that works?" Yes - you're talking to it.
Project Type: Internal Product / Production System Status: Live at blackskymedia.org Deployment: Railway (production), Local (development)