Cybersecurity Content Moderator

LLM-powered harmful content detection — text, image, and PDF

VIT Capstone | Hate Speech & Harmful Content Classification

Ollama LLM (WizardLM 7b / Gemma3 12b)
FAISS semantic search
Dockerised deployment
GitHub Actions CI/CD

AI cybersecurity system to detect hate speech and harmful content in text, images, and PDFs. Custom-trained Ollama model running locally with FAISS semantic retrieval, FastAPI backend, and Streamlit UI — packaged as a Docker image with auto-versioned CI push to Docker Hub.

What it does

A full-stack AI system that analyses user-submitted text, images, and PDF files for hate speech, harmful content, and cybersecurity threats. The backend runs a locally-hosted Ollama LLM (WizardLM 7b or Gemma3 12b) with FAISS-powered semantic search to retrieve similar past cases before generating a verdict.

Custom Ollama model FAISS vector search Text + Image + PDF GPU-accelerated v2.0.0 released

Python FastAPI Streamlit Ollama FAISS Sentence Transformers Docker GitHub Actions

Features

Text, Image & PDF Moderation: A single endpoint stack handles all three input types via POST /upload-text, /upload-image, and /upload-pdf.
AI-Powered Content Analysis: Custom-trained Ollama model (cyber-moderator-Wlm:7b / cyber-moderator-G3:12b) — WizardLM 7b for text, Gemma3 12b vision for images.
Semantic Search: FAISS index + Sentence Transformers retrieves the most similar prior cases to ground the LLM verdict.
User-Friendly UI: Streamlit frontend lets users upload files or paste text and see instant moderation results.
Containerised: Single Docker image, GPU-passthrough supported, available on Docker Hub.
Auto-versioned CI: GitHub Actions builds and pushes latest + semantic version tag on every merge to main.

API Endpoints

Endpoint	Method	Description
`/upload-text` or `/upload-image`	POST	Accepts plain text or image for moderation
`/upload-pdf`	POST	Accepts a PDF file — extracts text and analyses

Project structure

Content-moderator-image/
├── backend/
│   ├── backend.py          # FastAPI app — upload routes, FAISS, Ollama calls
│   ├── start.sh            # uvicorn launch
│   └── requirements.txt
├── frontend/
│   ├── frontend.py         # Streamlit UI
│   └── start_frontend.sh
├── models/
│   ├── Modelfile           # WizardLM 7b system prompt + params
│   └── Modelfile_vision    # Gemma3 12b vision config
├── images/                 # README screenshots
├── .github/workflows/
│   └── docker-build.yml    # GitHub Actions CI
├── Dockerfile
├── docker-compose.yml
└── start_services.sh       # Orchestrator: Ollama → backend → frontend

Architecture & system flows

End-to-end view: how a user input flows through Streamlit → FastAPI → Sentence Transformers + FAISS → Ollama LLM → verdict.

Full system — Streamlit FastAPI FAISS Ollama — **Full system architecture** User → Streamlit UI → FastAPI → Sentence Transformers + FAISS retrieval → Ollama LLM → verdict

Moderation pipeline input to verdict — **Moderation pipeline** Input → preprocess → embed → FAISS search → LLM classify → Harmful / Safe

Docker Compose service orchestration — **Docker Compose services** Ollama (port 11434, healthcheck) → backend (8000) → frontend (8501). Models persisted via volume.

GitHub Actions CI build push Docker Hub — **CI/CD** Push to main → GitHub Actions → auto-version → docker build → push latest + tag to Docker Hub

Custom Ollama model setup

Two fine-tuned models are created from base weights using Ollama Modelfile:

cyber-moderator-Wlm:7b — WizardLM 7b fine-tuned for text moderation (temperature 0.7, top_p 0.9)
cyber-moderator-G3:12b — Gemma3 12b vision model for image-based content classification

Modelfile (WizardLM)

FROM wizardlm2:7b
SYSTEM "Cybersecurity content moderation model"
PARAMETER "temperature" 0.7
PARAMETER "top_p" 0.9

Build & test the model

# Build
ollama create cyber-moderator-Wlm:7b -f /app/models/Modelfile

# Verify
ollama list

# GPU inference test
OLLAMA_USE_CUDA=1 ollama run cyber-moderator-Wlm:7b \
  "Analyze this message for harmful content."

FAISS semantic retrieval

Before calling the LLM, the backend embeds the input using Sentence Transformers and queries a FAISS index of past moderation examples. The top-k similar cases are injected as context into the LLM prompt, grounding the verdict in prior decisions and reducing hallucinations.

Supported moderation cases (capstone research)

Hate speech and slurs (text)
Weapons and drug imagery (images — WizardLM vision / Gemma3 12b)
Threat and harassment detection (text)
Harmful PDF content extraction and analysis
Toy / non-threatening object disambiguation (e.g., child with toy gun)

Backend processing output

FastAPI backend running — uvicorn output

FastAPI backend live — uvicorn on port 8000

Streamlit UI — text moderation

Streamlit text moderation interface

LLM verdict — harmful content flagged

Vision model — image moderation

Vision model detects real guns and drugs

Real weapons & drugs — flagged harmful

Vision model — child with toy gun classified safe

Child with toy gun — correctly classified safe

GitHub Actions CI build

GitHub Actions — build and versioning steps

Docker push — latest + semantic version tag

Quick start — Docker Hub image

# Pull the pre-built image
docker pull harishkumarthesde/content-moderator:latest

# Run with GPU support
docker run --gpus all \
  -p 8000:8000 -p 8501:8501 \
  --name content-moderator-container \
  harishkumarthesde/content-moderator:latest

# FastAPI:  http://localhost:8000
# Streamlit: http://localhost:8501

Docker Compose (full stack)

docker-compose up --build
# Starts: ollama (11434) → backend (8000) → frontend (8501)

docker-compose down

How Compose orchestrates services

ollama — ollama/ollama image, healthcheck on ollama list, persists models via volume.
backend — depends on ollama healthy; pulls base models, creates custom Modelfiles, starts uvicorn.
frontend — depends on backend; runs streamlit run frontend.py on port 8501.

GitHub Actions CI/CD

On every push to main: auto-calculates the next semantic version from git tags, builds the Docker image, pushes latest + version tag to Docker Hub, and creates a new git tag on the repository.

Trigger: push to main
  → Auto-version (e.g. v2.0.0 → v2.0.1)
  → docker build harishkumarthesde/content-moderator:latest
  → docker push :latest + :v2.0.1
  → git tag -a v2.0.1

Prerequisites to run

Docker or Podman installed
NVIDIA GPU + NVIDIA drivers + CUDA (for full performance)
Access to Docker Hub image harishkumarthesde/content-moderator:latest

Links

View on GitHub Docker Hub Ollama model

Back to Portfolio