AutoNAS - Neural Architecture Search with RL

AutoNAS is an intelligent Neural Architecture Search system that uses AI agents and reinforcement learning to design optimal neural networks automatically.

Comment

What is AutoNAS?

AutoNAS is an intelligent Neural Architecture Search system that uses AI agents and and reinforcement learning to design optimal neural networks automatically.

The Problem We Solve

Traditional NAS requires:

  • 🔥 Days/weeks of GPU training to evaluate each architecture
  • 💰 $10,000s in cloud compute costs Traditional NAS methods require expensive GPU training (hours to days)
  • 🧠 Deep expertise in neural network design
  • Manual architecture design is trial-and-error

Neural Architecture Search (NAS) Market Overview

  • Purpose: Enhances model performance, efficiency, and deployment adaptability across AI domains.
  • 2024 Market Size: Estimated between $420M – $670M globally.
  • Forecast (2033): Expected to reach ~$3.28B, driven by AutoML, edge AI, and enterprise AI adoption.
  • Growth Drivers:
    • Rising demand for efficient AI models.
    • Expansion of cloud-based AutoML platforms.
    • Edge and hardware-aware AI optimization.
    • Reduced time and cost for model development.

Our Solution

AutoNAS delivers:

  • Chat-based interface - describe your requirements in natural language
  • ⚡ Zero-cost evaluation - instant architecture scoring without training for quick check
  • Daytona sandbox training - for real validation
  • Online RL - learns and improves in real-time as you search
  • Complete automation - from user query to production-ready PyTorch code
  • 🤖 AI agents that understand your requirements

Three Core Innovations

1️⃣ Zero-Cost NAS (Instant Architecture Evaluation)

  • Uses NASWOT, Synflow, GradNorm proxies
  • Evaluates architectures in milliseconds (not hours)
  • Backed by NAS-Bench-201: 15,625 pre-computed architectures

2️⃣ Multi-Agent Orchestration

  • Strategic Agent: Understands natural language requirements
  • Research Agent: Finds state-of-the-art patterns
  • Generator Agent: Creates & evaluates candidates
  • Validation Agent: Ensures code correctness

3️⃣ Online Learning with Serverless RL

  • W&B Serverless RL trains policies in the cloud
  • Learns from every architecture evaluation
  • Gets smarter over time without local infrastructure

System Architecture

High-Level Design

┌─────────────────────────────────────────────────────────────┐
│                     Frontend (Next.js)                      │
│  Natural Language Query → Real-time Results & Visualizations│
└─────────────────────┬───────────────────────────────────────┘
                      │
              ┌───────▼───────┐
              │  FastAPI REST │
              │   + WebSocket │
              └───────┬───────┘
                      │
        ┌─────────────┼─────────────┐
        │             │             │
┌───────▼────────┐ ┌─▼──────────┐ ┌▼─────────────┐
│ Strategic Agent│ │ Zero-Cost  │ │ Serverless RL│
│  (Google ADK)  │ │    NAS/Daytona     │ │   (W&B)      │
└───────┬────────┘ └────────────┘ └──────────────┘
        │
   ┌────┴────┬─────────┐
   │         │         │
┌──▼──┐ ┌───▼───┐ ┌───▼──────┐
│Rsch │ │  Gen  │ │   Val    │
│Agent│ │ Agent │ │  Agent   │
└─────┘ └───────┘ └──────────┘

Design Rationale

  • Modular: Each agent is independent and specialized
  • Scalable: Serverless RL handles computation in the cloud
  • Observable: Weave traces every LLM call for debugging

Agent Orchestration & End-to-End Workflow

Complete User Journey

Step 1: User Input (Natural Language)

"Find me a CNN for CIFAR-10 with under 1M parameters"

Step 2: Strategic Agent (Google ADK + Gemini)

  • Parses natural language using Gemini 2.5 Flash
  • Extracts constraints: dataset=cifar10, max_params=1M
  • Coordinates the workflow

Step 3: Research Agent (Tavily API)

  • Searches academic papers and benchmarks
  • Finds SOTA patterns: ResNet blocks, SE modules, etc.
  • Returns design insights

Step 4: Generator Agent (LLM + RL)

  • LLM-based generation: Creates diverse candidates
  • RL-based generation: Uses learned policy for better suggestions
  • Zero-Cost evaluation: NASWOT + Synflow scores in <100ms
  • Logs to W&B: (state, action, reward) for online learning, Traces for LLM chat observability

Step 5: Results to User

  • Top-ranked architectures with explanations
  • Real-time learning curves showing RL improvement
  • Pareto frontier visualization (accuracy vs. efficiency)
  • Detailed evaluation metrics

Complete Tech Stack

🧠 AI & LLM Infrastructure

Technology Purpose Why We Use It
Google Cloud AI Multi-agent orchestration
→ ADK (Agent Dev Kit) Agent framework Structured agent design with built-in tools
→ Gemini 2.5 Flash LLM backbone Fast inference, strong reasoning

📊 ML Ops & Observability

Technology Purpose Why We Use It
Weights & Biases ML experiment tracking
→ Serverless RL Online policy learning PPO training without local GPU infra
→ Weave LLM call tracing Debug agent chains, cost tracking

🔍 Research & Search

Technology Purpose Why We Use It
Tavily API Research agent search Academic paper search, SOTA patterns

🛡️ Safe Code Execution

Technology Purpose Why We Use It
Daytona Validation agent sandbox Run generated code in isolated Python env

🏗️ Infrastructure

Technology Purpose Why We Use It
FastAPI Backend REST API Async support, WebSocket for real-time
Next.js 14 Frontend framework React Server Components, fast UX
TypeScript Type safety Catch errors at compile time

Built With

Share this project:

Updates