DjangoForge

Project Overview

DjangoForge takes a plain-language app description and turns it into a real, runnable Django project. The pipeline is labeled in its own architecture as NL → SIR → LLM Synthesis → Automated Evaluation → Live Preview / ZIP — natural language in, a structured intermediate representation in between, multi-file Django code out, automatically scored before it's returned.

Problem

Scaffolding a new Django app — models, views, URLs, admin registration, migrations, templates — is repetitive groundwork that has to happen before any real feature work can start. Generating that code with an LLM directly from a raw prompt is also unreliable on its own: without structure in between, output quality and consistency vary a lot. DjangoForge targets both problems at once.

Solution

A prompt box takes a description — typed freely, in English or Arabic, or picked from example prompts like "a task manager with projects and tasks" — and the backend converts it into a structured JSON schema (the SIR) before generating code. The result: full file structure, models, views, URLs, templates, migrations, and seed data, with a live in-browser preview and a one-click .zip download.

Key Features

• NL-to-SIR extraction: converts an unstructured prompt into a clean JSON schema defining models, fields, view types, and dynamic UI theme.
• Multilingual input: free-text prompts including non-English input — the demo shows a full app generated from an Arabic description.
• Multi-file LLM synthesis: generates full-stack Django code (models, views, URLs, templates, migrations) alongside a custom visual UI preview matching the prompt's domain.
• Automated quality evaluation: every generation is scored across five weighted dimensions before being handed back — see Engineering Highlights.
• Real-time SSE generation: built on FastAPI with Server-Sent Events for low-latency streaming instead of a single blocking request.
• File explorer: generated projects appear as a real file tree — manage.py, requirements.txt, app folder with models.py, views.py, urls.py, admin.py, migrations, and seed data.
• Code & Eval tabs: inspect generated source directly, alongside the evaluation score breakdown for that generation.
• Download: export the generated project as a .zip.

Architecture

A five-stage pipeline:

• Natural Language Input — the user's free-text description.
• SIR Extraction Engine — converts the prompt into a Structured Intermediate Representation (a normalized JSON schema of models, fields, and views).
• LLM Multi-File Synthesizer — generates models, views, URLs, templates, and a matching UI preview from the SIR.
• Automated Evaluation Engine — computes an Overall Quality Score (OQS) across five weighted dimensions.
• Live Preview & ZIP Output — renders the generated app in-browser and packages it for download.

The backend is a FastAPI service that calls the Claude API for generation, with Server-Sent Events streaming the response back to a lightweight static frontend.

Technology Stack

Engineering Highlights

• Structured Intermediate Representation: rather than generating code directly from a raw prompt, the pipeline passes through a structured JSON schema first — reducing inconsistency in the generated output and giving the synthesizer a well-defined contract to generate against.
• Five-dimension automated evaluation: every generated app is scored on Syntax Correctness (30%, via AST parsing across all Python modules), Field & Model Consistency (25%), View Coverage (20%), Template Completeness (15%), and URL Integrity (10%) — a real, if partial, stand-in for the manual code review a generated app would otherwise need.
• Streamed generation: built on FastAPI with Server-Sent Events, so multi-file generation streams back progressively instead of leaving the UI blocked on one long request.
• Rate-limited API surface: request throttling (SlowAPI) protects the underlying LLM API from abuse on a publicly reachable endpoint.
• Multilingual generation: the SIR extraction step accepts prompts in languages other than English, as demonstrated with Arabic input.

Results

In the demonstrated example, a single Arabic-language prompt ("I want a pink app for a flower agenda") produced a complete 19-file Django application — models, views, URLs, admin config, and seed data — with a working live preview and an Overall Quality Score of 97.3% from the automated evaluator. No broader benchmark numbers (average OQS across many prompts, generation latency at scale) are published here.

Links