Content Moderation API

AI-powered content moderation API built with FastAPI and DistilBERT. Sell API access to multiple companies — each gets isolated API keys, usage tracking, and flexible billing (token-based or monthly subscription).

📂 Project Structure

content_moderation/
├── api/
│   ├── main.py               # FastAPI app entry point, custom /docs (ReDoc)
│   ├── db.py                 # PostgreSQL async database layer
│   ├── models.py             # Pydantic request/response schemas
│   ├── middleware/
│   │   └── auth.py           # API key auth + quota enforcement
│   └── routes/
│       ├── health.py         # GET / (landing page) and /health
│       ├── admin.py          # Company registration, login, key management, billing
│       ├── moderate.py       # POST /moderate and /moderate/batch
│       └── superadmin.py     # Superadmin panel + contact form
├── src/
│   ├── filter.py             # Rule-based banned/political word filter
│   ├── moderate.py           # AI toxicity classification (DistilBERT)
│   └── train_transformer.py  # Model training script (class-weighted)
├── data/
│   ├── banned_words.txt      # Banned words list (English + Hindi)
│   ├── political_words.txt   # Political words list (global + India)
│   └── training_data.csv     # ~26,600 labeled examples (3 public datasets)
├── models/
│   └── transformer_model/    # Fine-tuned DistilBERT model files
├── static/
│   ├── index.html            # Landing page
│   ├── dashboard.html        # User dashboard panel
│   └── superadmin.html       # Superadmin panel
├── build_dataset.py          # Script to rebuild training_data.csv from HF datasets
├── .env                      # Environment variables
└── requirements.txt          # Python dependencies

⚙️ Prerequisites

Python 3.14+
PostgreSQL 17
Git

🚀 Setup & Installation

1. Clone the Repository

git clone <repository-url>
cd content_moderation

2. Create Virtual Environment

py -m venv venv
venv\Scripts\activate        # Windows
# source venv/bin/activate   # macOS/Linux

3. Install Dependencies

pip install -r requirements.txt

4. Setup PostgreSQL

Open pgAdmin or SQL Shell (psql) and run:

CREATE DATABASE content_moderation;

5. Configure Environment Variables

Edit the .env file in the project root:

DATABASE_URL=postgresql+asyncpg://postgres:YOUR_PASSWORD@localhost:5432/content_moderation

# Optional: set a fixed superadmin key (only used on first DB init)
# If blank, a random key is auto-generated and printed to console on first startup
SUPERADMIN_API_KEY=

# Superadmin web panel login credentials
SUPERADMIN_USERNAME=admin
SUPERADMIN_PASSWORD=softbender@123

If your password contains special characters (e.g. @), URL-encode them: @ → %40, # → %23, $ → %24

6. Train the Model (First Time Only)

python src/train_transformer.py

This saves the fine-tuned model to models/transformer_model/.

The training dataset (data/training_data.csv) contains ~26,600 labeled examples merged from 3 public Hugging Face datasets. To rebuild it from scratch, run python build_dataset.py.

7. Start the API Server

uvicorn api.main:app --host 0.0.0.0 --port 8000

Do not use --reload on Python 3.14 — it causes a multiprocessing spawn crash.

Server is now running at: http://localhost:8000

Tables are created automatically on first startup — no migrations needed. On first run, the SoftBenderAI superadmin API key is printed to the console. Save it immediately.

🧪 Demo Account (Ready to Test)

A demo company is seeded automatically on first startup. Use these credentials to test without registering:

Field	Value
Email	`demo@example.com`
Password	`password`
API Key	`cm_demo_key_for_testing_only_1234`
Billing	Token-based, 15,000 free tokens

Login via dashboard: http://localhost:8000/panel → enter email + password

Use API key directly:

curl -X POST http://localhost:8000/moderate \
  -H "X-API-Key: cm_demo_key_for_testing_only_1234" \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world, this is a test!"}'

🏢 SoftBenderAI — Superadmin Company

SoftBenderAI is the built-in superadmin company, seeded automatically on first startup.

No token limits, no daily limits — completely unrestricted
Used for platform management
API key is printed to console on first startup (or set via SUPERADMIN_API_KEY in .env)
All /superadmin/* endpoints require the SoftBenderAI API key

How to get the Superadmin key: Start the server and look for this in the console output:
============================================================
  SoftBenderAI Superadmin API Key: cm_xxxxxxxxxxxxxxxxxxxx
  Save this key — it won't be shown again!
============================================================
Or set SUPERADMIN_API_KEY=cm_yourkey in .env before first startup.

🌐 Web Panels

Browser-based UI — no curl needed.

Panel	URL	Who
Landing Page	http://localhost:8000	Public
User Dashboard	http://localhost:8000/panel	Any registered company
Superadmin Panel	http://localhost:8000/superadmin/panel	SoftBenderAI only
API Docs (ReDoc)	http://localhost:8000/docs	Anyone

User Dashboard (`/panel`)

Register — Create a new company account and get your API key
Login — Enter your email + password to access the dashboard
Overview — Token usage, request stats, API key list (paginated)
API Keys — Generate new keys, revoke existing ones
Billing — Buy token packs or subscribe to monthly plans
Test API — Moderate text directly from the browser

Superadmin Panel (`/superadmin/panel`)

Login with SUPERADMIN_USERNAME / SUPERADMIN_PASSWORD from .env (default: admin / softbender@123).

Overview — Platform-wide stats, revenue, top consumers
Companies — List, search, view details, activate/deactivate, delete
Manage — Set billing plans, grant tokens, generate keys for any company
Keys — Revoke any API key platform-wide
Plans & Pricing — View all plan tiers and token rates
Contacts — View contact form submissions with unread badge, mark as read

📖 API Documentation

Interface	URL
ReDoc Docs	http://localhost:8000/docs
Health Check	http://localhost:8000/health
OpenAPI JSON	http://localhost:8000/openapi.json

🔑 Authentication

All endpoints (except POST /admin/register, POST /admin/login, GET /, GET /health, POST /superadmin/contact) require an API key in the request header:

X-API-Key: cm_xxxxxxxxxxxxxxxx

💳 Billing Modes

Every company chooses a billing mode at registration. Both modes start with 15,000 free tokens.

Token-Based Billing

Pay as you go. Buy token packs whenever you need more.

Tier	Tokens	Price
Free (on signup)	15,000 tokens	₹0
Paid pack	15,000 tokens	₹300/pack

1 API request = 1 token
Free tokens never expire
Buy packs via POST /admin/topup
When tokens run out → 402 Payment Required

Monthly Subscription Plans

Fixed daily limits, billed monthly.

Plan	Daily Limit	Price/Month
`starter`	1,000 requests/day	₹499
`growth`	10,000 requests/day	₹1,499
`business`	50,000 requests/day	₹3,999
`enterprise`	Unlimited	₹9,999

Subscribe via POST /admin/plan/subscribe
Plan valid for 30 days from subscription date
Daily limit resets at midnight UTC
When daily limit is hit → 402 Payment Required
If billing_mode = monthly but no plan subscribed → falls back to token quota

📋 API Endpoints

Admin Endpoints

Register a Company

POST /admin/register
Content-Type: application/json

{
  "name": "Acme Corp",
  "email": "admin@acme.com",
  "password": "yourpassword",
  "billing_mode": "token"
}

billing_mode: "token" (default) or "monthly"

Response:

{
  "message": "Company registered successfully",
  "company_id": 1,
  "company_name": "Acme Corp",
  "billing_mode": "token",
  "api_key": "cm_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
  "free_tokens": 15000,
  "save_key": "Save this API key — it won't be shown again."
}

Save the api_key immediately — it is shown only once.

Login

POST /admin/login
Content-Type: application/json

{
  "email": "admin@acme.com",
  "password": "yourpassword"
}

Response:

{ "api_key": "cm_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" }

Login generates a new API key each time (keys are stored as SHA-256 hashes — irreversible).

Moderate a Single Text

POST /moderate
X-API-Key: cm_xxxxxxxx
Content-Type: application/json

{
  "text": "Your message here",
  "threshold": 0.7
}

Response:

{
  "allowed": true,
  "result": "approved",
  "blocked_reason": null,
  "confidence": 0.9821,
  "response_ms": 43,
  "free_tokens_remaining": 14957,
  "paid_tokens_remaining": 0
}

Batch Moderate (up to 50 texts)

POST /moderate/batch
X-API-Key: cm_xxxxxxxx
Content-Type: application/json

{
  "texts": ["Hello world", "I hate you", "Great product!"],
  "threshold": 0.7
}

Response:

{
  "results": [
    {"allowed": true,  "result": "approved", "blocked_reason": null,                     "confidence": 0.98, "response_ms": 12},
    {"allowed": false, "result": "blocked",  "blocked_reason": "Toxic/impolite language", "confidence": 0.91, "response_ms": 15},
    {"allowed": true,  "result": "approved", "blocked_reason": null,                     "confidence": 0.96, "response_ms": 11}
  ],
  "total": 3,
  "blocked_count": 1,
  "free_tokens_remaining": 14954,
  "paid_tokens_remaining": 0
}

Buy Token Pack

POST /admin/topup
X-API-Key: cm_xxxxxxxx
Content-Type: application/json

{ "packs": 2 }

Subscribe to Monthly Plan

POST /admin/plan/subscribe
X-API-Key: cm_xxxxxxxx
Content-Type: application/json

{ "plan": "growth" }

Generate Additional API Key

POST /admin/keys/generate
X-API-Key: cm_xxxxxxxx

Revoke an API Key

DELETE /admin/keys/{key_prefix}
X-API-Key: cm_xxxxxxxx

View Usage Stats

GET /admin/stats
X-API-Key: cm_xxxxxxxx

Superadmin Endpoints (SoftBenderAI only)

All /superadmin/* endpoints (except /superadmin/contact and /superadmin/login) require the SoftBenderAI API key.

Method	Endpoint	Description
`POST`	`/superadmin/login`	Login with username/password → returns API key
`GET`	`/superadmin/overview`	Platform-wide stats, revenue, top consumers
`GET`	`/superadmin/companies`	List all companies
`GET`	`/superadmin/companies/{id}`	Full stats for a specific company
`POST`	`/superadmin/companies/{id}/keys/generate`	Generate API key for any company
`POST`	`/superadmin/set-plan`	Set billing mode / plan / grant tokens
`DELETE`	`/superadmin/keys/{prefix}`	Revoke any API key
`PATCH`	`/superadmin/companies/{id}/toggle?activate=bool`	Activate or deactivate a company
`DELETE`	`/superadmin/companies/{id}`	Delete a company and all its data
`GET`	`/superadmin/plans`	List all plans and token pricing
`POST`	`/superadmin/contact`	Submit contact form (public, no auth)
`GET`	`/superadmin/contacts`	List all contact form submissions
`PATCH`	`/superadmin/contacts/{id}/read`	Mark a contact message as read

Set plan example:

POST /superadmin/set-plan
{
  "company_id": 5,
  "billing_mode": "monthly",
  "monthly_plan": "enterprise",
  "extra_tokens": 50000
}

🔧 Configuration

Adjust Toxicity Threshold

{ "text": "some text", "threshold": 0.5 }

Default is 0.7. Lower = stricter, higher = more lenient.

Add Banned Words

Edit data/banned_words.txt — one word/phrase per line. Supports English and Hindi. Leet-speak detection is built-in (e.g. f4ck → fuck).

Add Political Words

Edit data/political_words.txt — one word/phrase per line. Covers global leaders, Indian political parties, and policy terms.

Rebuild Training Dataset

python build_dataset.py

Downloads and merges 3 public Hugging Face datasets (~26,600 examples):

cardiffnlp/tweet_eval hate — ~13k English tweets
ucberkeley-dlab/measuring-hate-speech — ~6k comments
google/civil_comments — ~8k comments

Retrain the Model

python src/train_transformer.py

Uses class-weighted loss to handle imbalanced data (toxic ~27%, clean ~73%). Trains for 4 epochs with early stopping (patience=2).

🧪 Quick Test with curl

# 1. Use demo account (no registration needed)
curl -X POST http://localhost:8000/moderate \
  -H "X-API-Key: cm_demo_key_for_testing_only_1234" \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world"}'

# 2. Register a new company
curl -X POST http://localhost:8000/admin/register \
  -H "Content-Type: application/json" \
  -d '{"name": "Test Co", "email": "test@test.com", "password": "test123", "billing_mode": "token"}'

# 3. Login (get a new API key)
curl -X POST http://localhost:8000/admin/login \
  -H "Content-Type: application/json" \
  -d '{"email": "test@test.com", "password": "test123"}'

# 4. Batch moderate
curl -X POST http://localhost:8000/moderate/batch \
  -H "X-API-Key: cm_xxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{"texts": ["Hello!", "I hate you", "Great work"], "threshold": 0.7}'

# 5. Subscribe to monthly plan
curl -X POST http://localhost:8000/admin/plan/subscribe \
  -H "X-API-Key: cm_xxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{"plan": "growth"}'

# 6. Superadmin — list all companies
curl http://localhost:8000/superadmin/companies \
  -H "X-API-Key: <softbenderai_key>"

# 7. Superadmin — grant enterprise plan to company 5
curl -X POST http://localhost:8000/superadmin/set-plan \
  -H "X-API-Key: <softbenderai_key>" \
  -H "Content-Type: application/json" \
  -d '{"company_id": 5, "billing_mode": "monthly", "monthly_plan": "enterprise", "extra_tokens": 0}'

🤖 AI Moderation Pipeline

Every request goes through a 3-layer pipeline:

Empty check — blank text is blocked immediately
Rule-based filter (src/filter.py) — checks against banned_words.txt and political_words.txt with leet-speak normalization and substring scanning
AI classifier (src/moderate.py) — fine-tuned DistilBERT model scores toxicity probability; blocked if above threshold (default 0.7)

📊 Model & Dataset

Property	Value
Base model	`distilbert-base-uncased`
Training examples	~22,600 (train) + ~4,000 (test)
Toxic examples	~7,200
Clean examples	~19,400
Training epochs	4 (early stopping patience=2)
Loss function	Cross-entropy with inverse-frequency class weights
Optimizer	AdamW, lr=2e-5, weight_decay=0.01

🗄️ Database Schema

Table	Key Columns
`companies`	id, name, email, password_hash, is_superadmin, is_active, billing_mode, monthly_plan, plan_expires_at, free_tokens_used, paid_tokens_purchased, total_tokens_used
`api_keys`	id, company_id, key_hash, key_prefix, is_active, requests_today, total_requests, last_used, last_reset_date
`usage_logs`	id, company_id, api_key_prefix, text_length, result, blocked_reason, response_ms, is_free
`billing`	id, company_id, billing_type, tokens_purchased, packs, plan_name, amount_inr
`contact_messages`	id, name, email, phone, subject, message, is_read

All tables are created automatically on first startup via init_db(). Schema changes use ALTER TABLE ... ADD COLUMN IF NOT EXISTS migrations — no manual migration needed.

🔍 Troubleshooting

getaddrinfo failed on startup → PostgreSQL is not running. Start it from Windows Services or pgAdmin.

Invalid or inactive API key → Pass the key in X-API-Key header. Check that the company is active in the superadmin panel.

402 Token quota exhausted → Free 15,000 tokens used up. Buy more via POST /admin/topup or subscribe via POST /admin/plan/subscribe.

402 Daily limit reached → Monthly plan daily limit hit. Upgrade plan or wait for midnight UTC reset.

403 Superadmin access required → Calling a /superadmin/* endpoint without the SoftBenderAI API key.

CUDA out of memory → Reduce per_device_train_batch_size in src/train_transformer.py or run on CPU.

Server crashes with --reload → Known Python 3.14 + uvicorn issue. Run without --reload and restart manually after code changes.

SoftBenderAI key not saved → Drop the api_keys rows for admin@softbenderai.com in pgAdmin, restart the server — key will be regenerated and printed again.

Demo key not working → The demo account is seeded on first startup. If you started the server before this version, restart once — init_db() will seed it automatically (idempotent).

Name		Name	Last commit message	Last commit date
Latest commit History 5 Commits
api		api
data		data
src		src
static		static
.gitignore		.gitignore
README.md		README.md
build_dataset.py		build_dataset.py
content_moderation_linkedin_architecture_diagram.html.png		content_moderation_linkedin_architecture_diagram.html.png
content_moderation_linkedin_metrics_card.html.png		content_moderation_linkedin_metrics_card.html.png
content_moderation_linkedin_terminal_screenshot.html.png		content_moderation_linkedin_terminal_screenshot.html.png
generate_linkedin_assets.py		generate_linkedin_assets.py
linkedin_architecture_diagram.html		linkedin_architecture_diagram.html
linkedin_metrics_card.html		linkedin_metrics_card.html
linkedin_terminal_screenshot.html		linkedin_terminal_screenshot.html
requirements.txt		requirements.txt

Folders and files

Latest commit

History

Repository files navigation

Content Moderation API

📂 Project Structure

⚙️ Prerequisites

🚀 Setup & Installation

1. Clone the Repository

2. Create Virtual Environment

3. Install Dependencies

4. Setup PostgreSQL

5. Configure Environment Variables

6. Train the Model (First Time Only)

7. Start the API Server

🧪 Demo Account (Ready to Test)

🏢 SoftBenderAI — Superadmin Company

🌐 Web Panels

User Dashboard (/panel)

Superadmin Panel (/superadmin/panel)

📖 API Documentation

🔑 Authentication

💳 Billing Modes

Token-Based Billing

Monthly Subscription Plans

📋 API Endpoints

Admin Endpoints

Register a Company

Login

Moderate a Single Text

Batch Moderate (up to 50 texts)

Buy Token Pack

Subscribe to Monthly Plan

Generate Additional API Key

Revoke an API Key

View Usage Stats

Superadmin Endpoints (SoftBenderAI only)

🔧 Configuration

Adjust Toxicity Threshold

Add Banned Words

Add Political Words

Rebuild Training Dataset

Retrain the Model

🧪 Quick Test with curl

🤖 AI Moderation Pipeline

📊 Model & Dataset

🗄️ Database Schema

🔍 Troubleshooting

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages

User Dashboard (`/panel`)

Superadmin Panel (`/superadmin/panel`)