SAMM

• 

  home page

• 

  sketch selection

• 

  sketch

• 

  Final Design Render

• 

  Model Generation Closeup

• 

  Langsmith Tracing

💡 Inspiration

We realized product innovation inside large companies moves painfully slowly. An innovation manager with a strong idea still has to wait weeks for an agency to come back with research, then days to weeks for a concept, then more time for a deck, all before anyone in the room can even react to it. By the time the proposal lands, the moment has often passed. We built SAMM because the bottleneck isn't the ideas. It's the time and resources it takes to make one idea credible enough to present. We wanted to see what it looked like if that whole process collapsed into a single afternoon.

✨ What it does

SAMM compresses what normally takes an agency six to eight weeks into a single afternoon. You enter a company name, a product idea, and any internal documents you have. It runs brand intelligence and market research in parallel, builds a picture of your brand, maps competitors, and stress-tests the idea for fit and risk. Everything gates behind a human checkpoint before the design work starts. A five-agent swarm then works through naming, palette, materials, and strategy together, with every output scored against your brand before you see it. When you're done iterating, SAMM assembles the full stakeholder package.

The Pipeline

The Five-Agent Swarm

Two Loops, One Artifact

1. 🔄 Brand consistency loop: sits between the swarm and visual output, runs automatically, the user never touches it
2. 🔁 User iteration loop: the main product, where the innovation manager steers design direction while research and strategy stay locked underneath

🛠️ How we built it

The frontend is React and the backend is Python. The agent system runs on LangGraph, which lets all agents run in parallel so they are ready before the user hits the first checkpoint. Each agent calls the OpenAI API with scoped instructions and writes to a shared artifact that everything downstream reads from.

Architecture

React Frontend
      │
      ▼
Python Backend
      │
   ┌──┴───────────────────────────────────────┐
   │ LangGraph  — parallel agent execution    │
   │ OpenAI API — scoped per-agent calls      │
   │ Backboard  — RAG across agents & sessions│
   └──┬───────────────────────────────────────┘
      │
┌─────┴───────────────────────────────────────────┐
│ 🔍  Exa               Market + trend research   │
│ 🎬  Fal.ai (Composio) Video generation          │
│ 🧊  Meshy AI           3D mockup rendering      │
│ ☁️  Cloudinary         Visuals + video CDN      │
│ 🔗  Composio           Slack + Confluence       │
│ 📲  Pingram            SMS + email alerts       │
│ 🔐  Cystack            Vulnerability + secrets  │
│ 🧹  Polarity           PR review + code quality │
└─────────────────────────────────────────────────┘

Tech Stack

Cloudinary in SAMM

🏆 Accomplishments we're proud of

• 🏗️ A working end-to-end agent pipeline — from a company name and a rough idea to a pitch deck, product brief, video ad, 3D render, and market case in under four hours
• 🧩 Every single sponsor integrated meaningfully — each one made the system genuinely better
• 🧠 A shared artifact architecture — changing one thing only reruns what needs to rerun, research and strategy stay locked underneath
• 🎯 Something we could see actually seeing use — a real innovation team could open this on a Monday morning and walk into a meeting with something tangible by lunch

🚀 What's next for SAMM

The immediate next step is getting SAMM in front of real innovation teams at large companies and watching how they actually use it. We have assumptions baked into the flow that real sessions will either validate or break, and we want that feedback before we build more on top of it. On the product side, the biggest thing we want to tackle is proper multi-user sessions, where a whole team can work inside one artifact at the same time, with the strategist, the brand lead, and the innovation manager all steering the session together rather than one person running it and screenshotting the results for everyone else.

Built With

• backboard
• cloudinary
• composio
• cystack
• exa
• fal.ai
• langgraph
• meshyai
• openai
• polarity
• python
• rag
• react