• 

  Landing Dark Mode. The full Cyber-Forensic command center. Designed for deep dives where data visibility is paramount.

• 

  Landing Light Mode. A clean, minimal interface for rapid forensic investigation, focusing purely on the facts.

• 

  The Architecture of Truth. See how Market, Reddit, and Vision Scouts triangulate data to expose fake reviews and scams.

• 

  Toggle 'VS Mode' for a head-to-head deep dive, cross-referencing two products simultaneously.

• 

  Multimodal Evidence Locker. Visual Mode uses Gemini 3 Flash Vision to scan user photos or camera feeds for product data.

• 

  The Discovery Podium. A live, gamified feed of verified products and detected scam alerts currently under investigation.

SKEPTEK: The Truth Engine for E-Commerce

💡 Inspiration

We are living in a "Dead Internet" crisis.

• $770 Billion: The estimated global cost of fake reviews in 2025.
• 30-40%: The percentage of reviews on major marketplaces that are AI-generated or incentivized.
• The Problem: Shopping online has become a forensic investigation. Consumers are forced to open 50 tabs—Youtube, Reddit, Review sites—just to figure out if a $50 gadget is real or dropshipped junk.

We got tired of being victims. We realized that if bad actors are using AI to generate fake praise, we need better AI to detect the truth. We built Skeptek not as a shopping assistant, but as an Autonomous Forensic Team that works for you, not the seller.

🔎 What it does

Skeptek is an AI-powered "Truth Engine" that analyzes product links from any major retailer.

1. Market Scout: Tracks historical pricing to detect fake "80% off" deals.
2. Reddit Scout: Ingests entire subreddit communities using Gemini 3 Pro’s 1M context window to find authentic user sentiment (e.g., "battery died in 2 months") that is buried in comments.
3. Vision Scout: Uses Gemini 3 Flash Vision to watch YouTube reviews at 100x speed, frame-by-frame, detecting physical defects (wobbly hinges, cheap plastic) that the influencer might not mention.

It synthesizes this into a single Truth Score (0-100) and a verified "Verdict".

⚙️ How we built it

We built a "Split-Brain" Agentic Architecture:

• The Brain (Reasoning): Gemini 3 Pro handles the complex logic of cross-referencing conflicting data points (e.g., "The listing says 4K, but the Reddit user says 1080p").
• The Eyes (Vision): Gemini 3 Flash is our high-speed perception layer, processing extraction frames from video reviews.
• The Body (App):
  • Frontend: Next.js 14 (App Router) for a responsive, "Cyber-Forensic" UI.
  • Backend: Python (Flask) microservices for the heavy lifting—running selenium (undetected-chromedriver) to scrape live data and yt-dlp for video processing.
  • Orchestration: A custom TypeScript agent loop that delegates tasks to the Python backend via REST.
  • Database: Supabase (PostgreSQL) for storing "Field Reports" and caching product truth scores.

🚧 Challenges we ran into

• Anti-Bot Defenses: Major retailers have aggressive anti-scraping. We had to engineer a stealthy Python scraper using undetected-chromedriver to mimic human behavior just to get the raw HTML for Gemini to analyze.
• Video Context: Extracting meaningful frames from a 10-minute video without blowing up the token limit was tough. We solved this by using layout heuristic sampling (grabbing 1 frame every 5 seconds) and letting Gemini 3 Flash's speed handle the bulk throughput.
• Hallucinations vs. Sentiment: Early versions of the Reddit scout would get confused by sarcasm. We fixed this by using Gemini 3 Pro's larger context window to feed it entire thread histories, giving it the nuance to understand "Yeah, right, great product /s".

🏆 Accomplishments that we're proud of

• Autonomous Vision: Successfully building a pipeline that can "watch" a video and spot a physical defect is magical. Seeing Gemini flag a "scratch" on a pristine-looking product demo was a huge win.
• True Multimodality: We aren't just summarizing text. We are correlating pricing data, social sentiment, and visual evidence into one coherent verdict.
• Speed: Optimizing the agent loop so that a full forensic investigation takes seconds, not minutes.

🚀 What's next for Skeptek

• Browser Extension: Bringing the Truth Score directly onto the Amazon/Shopify page overlay.
• The "Memory Core": Building a shared, immutable ledger of product defects so that once Skeptek exposes a scam, it stays exposed forever.
• Mobile App: Scanning barcodes in physical stores to get an instant forensic report.

Built With

• gemini3flash
• gemini3pro
• geminiai
• googlegrounding
• logic)
• next.js
• python
• supabase
• tailwindcss
• typescript
• youtubeapi