---
title: AutoArchitect AI
emoji: 🧠
colorFrom: indigo
colorTo: purple
sdk: docker
pinned: true
license: mit
short_description: Plain English in. Trained AI agent out.
---

# AutoArchitect AI

**Build a classifier in seconds.** Image, text, tabular, audio, multimodal data. Plain English in. Trained AI agent out.

Not a chatbot or generator — we do classification, exceptionally well.

## What is this?

AutoArchitect AI is an autonomous AutoML system that takes a plain English problem description and produces a trained classification agent in seconds. No ML knowledge required.

### Verified on 5 domains
- **Image:** 86.6% on 67 unseen pothole images
- **Text:** 73.1% on 750 unseen fake news samples
- **Tabular:** 98% on synthetic fraud detection
- **Audio:** 100% on synthetic audio classification
- **Multimodal:** 87% on CIFAR-10 (CLIP zero-shot)

### How it works

1. **Brain Cores (3-stage thinking)** — Distilled from DeepSeek V3 ($0.22 total cost)
2. **ANAS Multi-Agent NAS** — Searches 266M architectures
3. **Foundation Models** — DINOv2, Wav2Vec2, CLIP, TabPFN
4. **Multi-agent Ensemble** — 12 specialized agents working in parallel

### Built by

**Dhanraj Atul Pandya**  
MS Computer Science Engineering  
Northeastern University Oakland  
GitHub: [@dhanraj176](https://github.com/dhanraj176)

### Tech stack

- Backend: Python 3.10, Flask, PyTorch, Transformers
- Brain: Qwen2.5-1.5B + 3 LoRA adapters (distilled from DeepSeek V3)
- Frontend: Vanilla HTML/CSS/JS
- Foundation models: DINOv2, Wav2Vec2, CLIP, XGBoost
- Storage: ChromaDB, semantic cache

### License

MIT