Security AI and Threat Intelligence

Overwatch AI

A high-performance AI command center intercepting multimodal scam threats in real time.

Overwatch AI is live as a security-focused command center built to expose AI-powered scams across voice, image and text channels. It delivers multimodal threat intelligence with a cinematic product experience and actionable verdict dashboards teams can trust under pressure.

Challenge

Scam attempts increasingly combine synthetic voice, manipulated visuals and engineered text pressure. Security teams needed a single system that can classify multimodal threats fast and explain risk with operational clarity.

Solution

We built a Next.js command center with a real-time scan pipeline powered by Gemini 3.1 Flash. The platform ingests audio, screenshots, text and URLs, then returns strict JSON intelligence including threat levels, authenticity scores and manipulation-tactic evidence.

Tech Stack and Architecture

Frontend and Experience

• Next.js 14 and React command-center architecture
• Tailwind CSS and Framer Motion for cinematic interaction design
• State-driven UI flow: Idle, Scanning, Verdict
• Responsive glassmorphism styling with threat-color accents

AI and Detection Engine

• Gemini 3.1 Flash multimodal models via @google/generative-ai
• Dynamic system prompt construction by media type
• Strict JSON schema enforcement for reliable machine-readable verdicts
• Threat level, authenticity scoring and tactic-level intelligence outputs

Backend and Delivery

• Next.js serverless scan API for secure inference orchestration
• Payload parsing for text, URLs and file uploads
• Production deployment with high-speed global delivery

Engineering Phases

Phase 1: Threat Intelligence Product Definition

Defined the command-center UX around rapid triage: from input ingestion to final verdict in a single guided interface.

Phase 2: Multimodal Scan Pipeline

Implemented the scan API to process voice notes, screenshots, text and URLs, then route each input through media-aware AI prompting paths.

Phase 3: Structured Verdict Intelligence

Added strict structured-response contracts so every analysis returns consistent fields for threat level, authenticity and manipulation techniques.

Phase 4: Cinematic Operational Interface

Delivered a high-feedback interface with scanning animations, radar pulses and final verdict dashboards that improve confidence and speed during investigations.

Debugging Highlights

• Stabilized multimodal payload handling to support mixed input types without scan interruptions.
• Hardened JSON schema validation to prevent malformed AI outputs from reaching the verdict UI.
• Optimized animated scanning states for responsive performance across desktop and mobile devices.

Roadmap

• Expand threat taxonomy and intelligence memory for broader scam pattern coverage.
• Add investigator collaboration modes and historical case comparison workflows.