Hero Banner

Top Builders

Explore the top contributors showcasing the highest number of app submissions within our community.

Meta

Meta, founded in 2004, is a global technology leader that revolutionizes how people connect and interact in the digital world. Originally known as Facebook, Meta is renowned for its pioneering advancements in social media, with platforms like Facebook, Instagram, and WhatsApp, which collectively reach billions of users worldwide. In addition to its social media prowess, Meta is a global technology company at the forefront of AI innovation, focusing on enhancing human connectivity and creating immersive digital experiences. Among its leading products related to AI technology are the LLaMA (Large Language Model Meta AI) series and Meta AI.

General
CompanyMeta Platforms, Inc.
FoundedJanuary 4, 2004
HeadquartersMenlo Park, California, U.S.
Repositoryhttps://github.com/facebook

Key Products and Research

Meta has developed a range of AI products designed to enhance various aspects of technology and user experience. Here’s a brief overview of these AI products:

LLaMA (Large Language Model Meta AI)

LLaMA is a series of large language models designed for natural language processing tasks. These models, including the latest LLaMA 3.1, are known for their advanced capabilities in text generation, understanding, and multilingual processing. They are available as open-source models, promoting innovation and research in AI​ Meta | Social Metaverse Company,Facebook.

Meta AI

Meta AI is an intelligent assistant integrated across Meta’s platforms, such as Facebook, Instagram, WhatsApp, and Messenger. Powered by LLaMA models, it helps users with tasks like content creation, information retrieval, and personalized interactions Meta | Social Metaverse Company.

PyTorch

PyTorch is an open-source machine learning library developed by Meta and widely used in both research and industry. It provides tools for building and training deep learning models and has become a standard framework in the AI community​ Facebook.

Meta AI Research (FAIR)

Meta’s AI research division, formerly known as FAIR (Facebook AI Research), focuses on advancing the field of AI through open research and collaboration. This division works on various AI challenges, including computer vision, natural language processing, and generative AI​ Facebook.

Meta AI in the Metaverse

Meta is also incorporating AI into its metaverse initiatives, using AI to create immersive experiences in virtual and augmented reality. This includes developing AI-driven avatars, enhancing virtual environments, and improving interaction within the metaverse​ Meta | Social Metaverse Company.

AI for Ads

Meta leverages AI to optimize ad targeting, delivery, and measurement across its platforms. AI algorithms analyze vast amounts of data to improve the effectiveness of advertising campaigns, making them more relevant to users and efficient for advertisers​ Meta | Social Metaverse Company.

LLaMA Impact Grants

The LLaMA Impact Grants program, launched by Meta, aims to support and encourage the innovative use of its LLaMA (Large Language Model Meta AI) models to address critical challenges in various sectors, including education, environmental sustainability, and public good. This initiative offers financial grants and resources to researchers, nonprofits, and other organizations that seek to leverage LLaMA models for impactful projects. The program highlights Meta’s commitment to responsible AI development and its belief in the potential of AI to drive positive social change.

For more details, visit the LLaMA Impact Grants page.

Edit on GitHub

Meta AI Technologies Hackathon projects

Discover innovative solutions crafted with Meta AI Technologies, developed by our community members during our engaging hackathons.

Compliance Copilot - AI Transaction Alert Triage

Compliance Copilot - AI Transaction Alert Triage

Compliance Copilot addresses one of the biggest challenges in financial crime operations: alert fatigue caused by high false-positive rates in transaction monitoring systems. Compliance teams often receive thousands of alerts weekly, making it difficult to quickly identify truly suspicious behavior. This project introduces an AI-powered alert triage system that combines deterministic, auditable risk scoring with generative AI explanations. Each transaction alert is evaluated using transparent risk signals such as transaction amount deviation, behavioral patterns (velocity and burst activity), and geographic corridor risk. Alerts are scored on a 0–100 scale and classified as High, Medium, or Low risk. A local large language model is then used to generate concise, analyst-ready explanations that clearly summarize why an alert was flagged and recommend next investigative steps. This approach preserves explainability and regulatory trust while significantly reducing manual review effort. By collapsing large alert volumes into a small set of high-confidence cases with clear reasoning, Compliance Copilot demonstrates how AI can improve operational efficiency, decision quality, and regulatory readiness in real-world compliance workflows. All data used is synthetic and the system is designed to integrate easily into existing monitoring pipelines.

Deriv Sentinel -Self-Healing AI WAF for LLM Agents

Deriv Sentinel -Self-Healing AI WAF for LLM Agents

Deriv Sentinel is an AI-powered Web Application Firewall that protects LLM agents from prompt injection and data leakage through a continuous red-team-and-heal cycle. The Problem: Traditional WAFs can't protect AI agents. Prompt injection is the SQL injection of the AI era - natural language attacks bypass conventional input validation, and patching one technique just leads attackers to find new ones. Our Solution: Instead of waiting for attacks, Deriv Sentinel attacks itself first, then autonomously patches the vulnerabilities it discovers. How It Works: 1. Attack — An attacker model generates realistic social engineering prompts enriched with Shadow RAG context (fake internal documents as honeypots). 2. Defend — Bastion (llama3.1:8b), our protected LLM loaded with simulated internal data, responds to each attack. 3. Audit — ShieldGemma (shieldgemma:2b) audits every response for data leakage and policy violations, backed by deterministic pattern matching as a second detection layer. 4. Heal — When a breach is detected, the Heal Engine injects a vaccine guardrail and redacts the exploited knowledge section. The same attack now gets blocked — without retraining. 5. Human-in-the-Loop — Analysts can approve/reject heals or enable auto-heal for autonomous defense. Key Innovations: - Knowledge Base Redaction — We remove leaked data from context entirely. LLMs can't leak what they don't have. - Multi-Layered Defense — AI auditor + deterministic matching + post-processing enforcement. - Instant, Reversible Fixes — Runtime prompt patches. No fine-tuning, no redeployment. - Adaptive — Each breach teaches the system a new defense. Demo: Reset → Run red-team → Bastion leaks secrets → ShieldGemma detects → Heal applied → Same attack blocked. Self-healing proven in five minutes.

Drishpex

Drishpex

Fraud detection today is broken. Rule-based systems generate 90-95% false positive rates, analysts waste hours investigating noise, and by the time someone writes a new rule, fraudsters have already moved on. Interpol estimates we catch just 2% of global financial crime. We built a fraud detection agent that actually closes the loop. Transactions stream in and get scored in under 100ms by an XGBoost model trained on 35 behavioral features. High-risk transactions automatically become cases. An LLM powered by Llama 3.1 investigates each case and produces a structured report — turning a 20-minute manual investigation into a 5-second read. When an analyst confirms a label, the model retrains itself in the background. No data scientist needed. Meanwhile, a graph mining engine runs autonomously on the transaction network using Tarjan's SCC, HITS, and sliding-window algorithms to discover wash trading rings, hub accounts, and velocity clusters. Those patterns flow back into the scorer as new features — the system literally gets smarter on its own. The results speak for themselves: false positive rate dropped from 95% to 4.3%, model F1 improved from 0.57 to 0.967, and detection latency went from hours to milliseconds. The analyst does two things — read and label. Everything else is the agent. Built with FastAPI, XGBoost, NetworkX, Ollama, and Streamlit. Fully dockerized, deployed on AWS, and running live right now.

Deriv Sentinel AI - Autonomous Security Platform

Deriv Sentinel AI - Autonomous Security Platform

Deriv Sentinel AI is a full-stack AI security platform (NestJS + Next.js + Llama 3.2 via Ollama) with three autonomous agents that form a self-improving defence loop. AI WAF — Multi-layered prompt firewall that classifies inputs as safe/suspicious/malicious, validates AI outputs for prompt leakage and PII, and adapts over time by learning from red team bypass patterns. AI Pentester — Runs 7 parallel security checks against any target: exposed .env files, outdated dependencies, Next.js CVE detection, missing security headers, CORS misconfiguration, information disclosure, and sensitive directory exposure. The LLM generates severity ratings and remediation steps for each finding. AI Red Team — Generates adversarial attacks across 7 categories (prompt injection, jailbreaks, encoding evasion, social engineering, etc.), fires them against the WAF, and records which ones bypass detection. The key innovation is the closed feedback loop: red team bypasses are automatically injected into the WAF's classification prompt, patching blind spots without human intervention. Every attack run makes the WAF stronger. All results persist to SQLite with a real-time dashboard featuring charts, activity feeds, and cross-module stats. An AI chatbot demonstrates the defence layers in a conversational context — every message passes through both input and output WAF validation.

RouterAI

RouterAI

RouterAI is an intelligent AI API gateway that removes the complexity and cost inefficiencies of using multiple AI providers. Instead of manually choosing between OpenAI, Anthropic, Google, or Hugging Face, developers send every request to a single RouterAI API. RouterAI automatically analyzes the prompt, understands the task type, and routes it to the most suitable and cost-effective AI model. For simple tasks, RouterAI can use free or low-cost models like Llama. For advanced reasoning or coding, it selects powerful models like GPT-4, Claude, or Gemini. This smart routing helps developers avoid overpaying and reduces AI API costs by 30–70% without sacrificing quality. RouterAI introduces pay-per-request USDC micropayments on blockchain, so users only pay for what they actually use. Each request triggers a tiny USDC payment in real time, enabling sub-cent pricing that traditional payment systems cannot support. This makes RouterAI ideal for developers, startups, and autonomous AI agents that need flexible, transparent, and scalable billing. In simple terms: “One API, all AI models, pay only for what you use with USDC.” Technically, RouterAI works as follows: developers send an OpenAI-compatible API request, Gemini AI classifies the task, RouterAI selects the optimal provider, processes the USDC payment, and returns the AI response along with the exact cost deducted from the user’s wallet. RouterAI simplifies AI integration, cuts costs, and enables the future of agent-driven commerce.

Arcfeed

Arcfeed

Adds a demo route (/api/demo/402-flow) and an info endpoint to illustrate an HTTP 402 payment flow in-app. A first request (no X-Payment-Tx) returns a 402 JSON with payment.instructions, receiver, amount, and a demoTxHash; repeating the request with X-Payment-Tx: demo-payment-0x1234567890abcdef returns mock Aave yields data and paymentVerified: true. Includes a runnable helper script demo-402.sh that defaults to HTTPS, follows redirects, accepts an optional API_KEY env var (X-API-Key), and pretty-prints JSON when jq is available — use HOST and API_KEY to point at a deployed host. Intended for live demos and documentation to show the exact UX of requiring payment proof, how to present the 402 payload, and the transition to successful access. Note: ensure the backend API_KEY_SECRET matches the frontend VITE_API_KEY (or supply API_KEY when running the script).