MuLM: No-Code Conversational AI System Builder

The Problem

Building conversational agents requires weeks of coding for model integration, routing logic, and API orchestration. Developers waste 80% of their time on infrastructure instead of innovation. Expensive cloud APIs drain budgets and compromise privacy, while platform lock-in limits flexibility.

Our Solution

MuLM (Multi-Model Language Machine) transforms natural language descriptions into fully functional conversational AI agents with ElevenLabs voice integration—ready to test and export in minutes.

How It Works

Describe Your Agent: Type “Build a customer support agent that detects angry customers, escalates negative cases, and responds to others with a warm voice.”

AI Generates Workflow: In 15 seconds, creates visual workflow: input → sentiment analysis → router → response → ElevenLabs voice. Each node has generated code you can inspect.

Test Instantly: Enter sample inputs, watch real-time execution, hear voice responses. Unlimited testing without deployment or costs.

Customize: Click nodes to modify parameters, routing logic, voice characteristics, or edit Python code directly.

Export: Download as Jupyter Notebook or Python package ready to deploy on your infrastructure.

Key Innovations

Privacy-First: Self-contained systems run on your infrastructure. Data never leaves your environment. Own the complete code.

Intelligent Selection: Chooses between lightweight models, transformers, rule-based logic, and ElevenLabs services based on requirements.

Deep ElevenLabs Integration: STT for voice input, TTS for output with emotion/accent control, 32 languages, seamless workflow integration.

Demo Use Cases

Customer Support: Sentiment routing escalates angry customers, generates empathetic responses for others with warm voice.

Medical Triage: Extracts symptoms, classifies urgency, recommends action with professional voice. “Severe chest pain” → HIGH urgency → “Seek emergency care.”

Multilingual Sales: Auto-detects language, responds with matching accent. Spanish → Latin accent. Hindi → Indian accent.

Technical Stack

• AI Generation: Claude API converts natural language to workflows
• Visual Builder: React Flow for node-based editing
• Code Editor: Monaco Editor for customization
• Voice: ElevenLabs TTS/STT, 32 languages, emotion control
• Models: HuggingFace transformers, custom models, scikit-learn
• Export: Jupyter Notebook and Python package formats

Real-World Impact

Researchers: Test architectures without infrastructure. Export notebooks for reproducible research. Prototype in minutes.

Startups: Build specialized agents without ML teams. No vendor lock-in. One-time build vs. recurring API costs.

Enterprises: Privacy-compliant (on-premise). Customizable. Deploy anywhere.

Developers: Focus on problems, not infrastructure. Learn from generated code. Iterate 10x faster.

Goals

Hackathon: Demonstrate prompt → generate → test → export workflow. Showcase 3 agents. Prove ElevenLabs value.

Post-Hackathon: Community model library, bring-your-own-model, advanced architectures, template marketplace, enterprise features.

Contribute

Now: Frontend (UX), backend (optimization), ML (testing), voice (ElevenLabs patterns).

Later: Open-source release. Contribute models, templates, documentation, integrations, translations.

Why It Matters

Traditional development: 5-7 weeks. With MuLM: 8 minutes. 500x faster.

Makes voice AI accessible to solo developers, researchers, domain experts, students. Privacy-first enables GDPR/HIPAA compliance. No recurring costs after export.

Stack: React, FastAPI, HuggingFace, ElevenLabs, Claude, PostgreSQL

Team: Subhasis Jena, Brunda N. (Team Maelstrom)

Contact: [email protected]

Making conversational AI accessible and privacy-respecting.