AWS Builder Center
AIdeas Finalist: ASET – Academic Safety and Evidencing Truth

AIdeas Finalist: ASET – Academic Safety and Evidencing Truth

ASET (Academic Safety and Evidencing Truth) is a scientific claim verification engine built for the AI era — where information can be generated instantly, but trust cannot, to address the growing problem of hallucinated citations and AI-generated information. Instead of probabilistic prediction, ASET retrieves, analyzes, and synthesizes evidence from 1.2M+ indexed research papers to produce literature-backed verification reports for claims extracted from text, documents, webpages, images, and videos.

Security-Focused Systems Builder
App Category: Social Impact
Team: ZeTech
  • Om Singh — jayom5797@
  • Utsav Singh — utsavsingh35@
  • Vikas Tiwari — vikas2731@



My Vision

When we submitted ASET to the AWS 10,000 AIdeas competition, we had built a space science claim verifier. The judges called it out directly: limited to one domain, no differentiation from Semantic Scholar, no user validation.
They were right. So we rebuilt it entirely.
ASET v2 is not a search engine for papers. It is a verification pipeline for content you are already consuming — videos, documents, lectures, social media posts. That is a fundamentally different product from anything that exists today.
The problem has not changed: 46% of AI-generated academic citations are fabricated. Students submit papers with fake references. Teachers cannot fact-check fast enough. Content creators spread misinformation they genuinely believe is true. The crisis is real and growing.
What changed is our understanding of who actually needs this and how they use it. ASET now serves students, teachers, journalists, researchers, and content creators — anyone who needs to know: is this claim backed by real science?



Why This Matters

Semantic Scholar, Elicit, and Consensus are tools for researchers who want to find papers. Their user is someone sitting down to do a literature review.
ASET's user is a student watching a YouTube lecture at midnight, a journalist reading a PDF report, a teacher reviewing a student submission. They are not looking for papers. They have content in hand and need to know: is this true?
That is not the same product. That is not the same workflow. And no existing tool solves it.
ASET v2 addresses the judges' core criticisms directly:
"Limited to space science only" → Now 8 domains, 1.2M+ papers:
  • Space Science (972K, original)
  • Biology, Medicine (PubMed)
  • Chemistry, Physics, CS, Engineering (arXiv OAI-PMH)
"No user validation" → Real users from IIT Bombay, AIIMS, TIFR, IISc, ISRO onboarded and tested across all three modes.
"Insufficient differentiation" → Three verification modes no competitor offers, plus a self-growing database.



How I Built This

The Three Modes

Mode 1 — Single Claim: Type any scientific claim, get a trust score with supporting papers in under 200ms.
Mode 2 — YouTube: Paste any YouTube URL. ASET extracts the transcript, breaks it into atomic factual claims using Groq LLaMA 3.3 70B, and verifies each one against peer-reviewed research. A student watching a health lecture gets a full verification report before they cite anything.
Mode 3 — Document: Upload a PDF, DOCX, or image. ASET extracts text (with Tesseract OCR for images), identifies all verifiable claims, and returns a claim-by-claim report with an overall trust score.

Architecture

React Frontend (Vite)
AWS CloudFront + S3 Node.js Backend (Express)
aset-ai.tech
│ │
└──────── HTTPS API ────────────┘
│
Turso (libSQL/SQLite + FTS5)
1.2M papers, 72 topics
│
Groq LLaMA 3.3 70B
Claim extraction + verification
│
arXiv OAI-PMH + PubMed
Self-growing database


The Verification Pipeline

Every claim — whether typed, from a document, or from a YouTube transcript — goes through the same pipeline:
  1. Domain Detection: 28 keyword patterns route the claim to the correct scientific domain
  1. FTS5 Search: Pre-indexed 1.2M papers searched in 50-200ms
  1. Self-Healing: If no local papers found, ASET fetches from arXiv + PubMed in real time, permanently stores them, and the database grows with every query
  1. AI Verification: Groq LLaMA 3.3 70B analyzes each paper for stance (supports/contradicts/neutral), confidence score, and evidence
  1. Trust Score: Weighted aggregate across all papers → verdict + key findings


Why FTS5 Over Vector Search

Sub-200ms search is only possible because everything is pre-indexed locally. Hitting an external API at query time adds 300-800ms per claim. In batch mode (20+ claims from a PDF), that becomes 6-40 seconds. Our pre-indexed FTS5 architecture wins on speed. This is our technical moat.



Why EC2 + Nginx Over Fully Managed Deployment

As ASET expanded into multi-mode verification workflows and browser-integrated APIs, we shifted toward a lightweight EC2-based deployment architecture for greater runtime control and operational flexibility.
The backend now runs on AWS EC2 t2.micro with PM2 process management, Nginx reverse proxying, and HTTPS secured through Let’s Encrypt. This architecture reduced operational complexity while remaining within AWS free-tier limits during the current growth phase.
As ASET scales further, the infrastructure can later transition toward more distributed and auto-scaling deployment patterns.


Self-Growing Database

When FTS5 returns zero results, the backend calls  fetchAndStorePapers()  — queries arXiv and PubMed in parallel, classifies each paper across 28 domain patterns using best-score matching, inserts into Turso, updates the FTS index. Future queries find these papers immediately. The database grows with every unverifiable claim.



AWS Infrastructure Deep Dive

ASET’s core application infrastructure is deployed on AWS using a lightweight, cost-efficient architecture optimized for real-time verification workloads.
The frontend is distributed globally through CloudFront, while the backend API runs on EC2 with Nginx and PM2. The research corpus is indexed using Turso/libSQL with SQLite FTS5 for high-speed lexical retrieval.

CloudFront + Custom Domain (aset-ai.tech)

The frontend is served through CloudFront with a custom domain registered via Route 53 and secured with an ACM SSL certificate.
  • SSL: ACM certificate (us-east-1), TLSv1.2_2021 minimum
  • Edge caching: React bundle cached at 450+ edge locations globally
  • Cache hit ratio: 85%+ — most users never hit the origin
  • Compression: Gzip enabled — 2MB bundle → 600KB over the wire
  • Invalidation: Every deployment triggers  /*  invalidation for instant updates

S3 Static Hosting

The React frontend (built with Vite) is deployed to S3 with static website hosting.
  • Contents: index.html, JS bundle, CSS, fonts, videos, SVGs
  • SPA routing: Error document set to index.html for client-side routing
  • Public access: Enabled via bucket policy for CloudFront origin
  • Deployment:  aws s3 sync --delete  for atomic deployments



ECR (Elastic Container Registry)

Docker images are stored in ECR and pulled by App Runner on each deployment.
  • Repository: aset-backend (us-east-1)
  • Image: Node 22 Alpine + Express + all dependencies
  • Tagging:  latest  tag, immutable deployments
  • Security: IAM role  AppRunnerECRRole  with  AWSAppRunnerServicePolicyForECRAccess 

SES (Simple Email Service)

Password reset OTP emails are sent via AWS SES from the App Runner backend.
  • Use case: 6-digit OTP emails for password reset flow
  • Integration: Nodemailer with SES SMTP in the Node.js backend
  • Template: Custom HTML email with ASET branding

IAM (Identity and Access Management)

Least-privilege IAM roles secure every service-to-service interaction.
  • AppRunnerECRRole: Allows App Runner to pull images from ECR
  • Policy:  AWSAppRunnerServicePolicyForECRAccess 
  • om07 user: Admin access for CLI deployments
  • No hardcoded credentials: All AWS SDK calls use IAM roles, not access keys in code

Turso (libSQL — SQLite-compatible edge database)

While Turso is not an AWS service, it runs on AWS infrastructure (ap-south-1) and is the database layer for ASET's 1.2M paper index.
  • Engine: libSQL (SQLite-compatible) with FTS5 full-text search
  • Size: 1.2M papers, 72 topics, 28 subtopic categories
  • Query time: 50-200ms for FTS5 full-text search
  • Auth: JWT token authentication

Complete AWS Cost Breakdown

The entire production stack costs is now free
Service
Usage
Monthly Cost
CloudFront
~50GB transfer
$0 (free tier)
S3
50MB storage
$0 (free tier)
EC2 t2.micro
750 hours
$0 (free tier, 12 months)
ACM
SSL certificate
$0 (always free)
SES
<100 emails/day
$0 (free tier)
Total

$0/month

Browser Extension

ASET is now available as a browser extension, allowing users to verify scientific and factual claims directly on webpages without leaving their workflow.

Features

  • Highlight any text on a webpage
  • Right-click → “Verify with ASET”
  • Slide-in verification panel with:
  • trust score
  • supporting papers
  • evidence summary
  • Draggable and dismissible interface
  • No login required

Installation

  • Available inside the  aset-extension  folder of the GitHub repository
  • Load unpacked in Chrome or Edge developer mode
  • Microsoft Edge Add-ons release pending review

Why It Matters

The browser extension transforms ASET from a standalone verification website into a browser-native trust layer for online information consumption.


╔═══════════════════════════════════════════════════════════════════════════════╗
║ ASET — Complete AWS Infrastructure ║
╚═══════════════════════════════════════════════════════════════════════════════╝

┌─────────────────┐
│ User Browser │
│ aset-ai.com │
└────────┬────────┘
│ HTTPS
▼
┌────────────────────────────────────────────────────┐
│ AWS Route 53 │
│ aset-ai.com ──► CloudFront Alias Record │
│ www.aset-ai.com ──► CloudFront Alias Record │
└────────────────────────┬───────────────────────────┘
│
▼
┌────────────────────────────────────────────────────┐
│ AWS CloudFront (E38WWI7NAQX84D) │
│ • 450+ global edge locations │
│ • ACM SSL Certificate (aset-ai.com) │
│ • TLSv1.2_2021 • Gzip compression │
│ • Cache hit ratio: 85%+ │
└──────────┬─────────────────────────┬──────────────┘
│ │
┌──────────▼──────────┐ ┌──────────▼──────────────┐
│ AWS S3 │ │ AWS App Runner │
│ Static Frontend │ │ Node.js Backend │
│ │ │ │
│ • React + Vite │ │ • 1 vCPU / 2GB RAM │
│ • index.html │ │ • Auto-scales to zero │
│ • JS/CSS bundles │ │ • HTTPS auto-managed │
│ • Fonts & videos │ │ • Health: /health │
│ • SPA routing │ │ • Env vars injected │
└─────────────────────┘ └──────────┬──────────────┘
│
┌──────────────────────────┼──────────────────────────┐
│ │ │
┌─────────▼──────────┐ ┌──────────▼──────────┐ ┌──────────▼──────────┐
│ AWS ECR │ │ Turso DB │ │ AWS SES │
│ Container Registry│ │ (libSQL on AWS) │ │ Email Service │
│ │ │ │ │ │
│ • Docker image │ │ • 1.2M papers │ │ • OTP emails │
│ • Node 22 Alpine │ │ • FTS5 search │ │ • Password reset │
│ • App Runner pull │ │ • 72 topics │ │ • Verified sender │
│ • IAM role auth │ │ • 50-200ms queries │ │ • HTML templates │
└────────────────────┘ └──────────────────────┘ └─────────────────────┘

┌────────────────────────────────────────────────────────────────────────┐
│ AWS IAM │
│ AppRunnerECRRole ──► AWSAppRunnerServicePolicyForECRAccess │
│ om07 user ──► AdministratorAccess (CLI deployments) │
└────────────────────────────────────────────────────────────────────────┘

┌────────────────────────────────────────────────────────────────────────┐
│ AWS ACM │
│ Certificate: aset-ai.com + www.aset-ai.com │
│ Validation: DNS via Route 53 CNAME • Auto-renewal • Free │
└────────────────────────────────────────────────────────────────────────┘

═══════════════════════════════════════════════════════════════════════════════
Request Flow: User → Route 53 → CloudFront → S3 (static) / App Runner (API)
Data Flow: App Runner → Turso (papers) → Groq (AI) → arXiv/PubMed (self-heal)
Auth Flow: App Runner → SES (OTP email) → User → JWT token
═══════════════════════════════════════════════════════════════════════════════




Demo


🚀Live app:  https://aset-ai.tech 
The demo shows:
  1. Single claim: "The amygdala controls emotional memory" → Supported 85%, 3 supporting papers
  1. Document upload: 7-page biology study guide → 13 claims, 11 verified, 2 unverifiable
  1. YouTube mode: Science education video → transcript extracted, 9 claims verified with trust scores
  1. Paper Search: Search 1.2M papers by title/author/keyword — no login required



Current Capabilities

  • Verification against 1.2M+ indexed research papers
  • Multi-format inputs:
  • text
  • PDF
  • DOCX
  • images
  • YouTube transcripts
  • Browser extension for inline webpage verification
  • Real-time semantic verification workflows
  • Claim-by-claim document analysis
  • Evidence-backed trust scoring
  • Self-expanding research corpus ingestion



What I Learned

The judge feedback that resonated most: "Insufficient differentiation from established tools."
This forced us to ask the right question: what do Semantic Scholar, Elicit, and Consensus actually do? They help researchers find papers. They do not help students verify the content they are already consuming. That gap is where ASET lives.
The three-mode architecture is our answer. It is not a feature list. It is a different use case entirely.
Key technical insights:
FTS5 with pre-indexed data beats every API-based approach for latency. We tested arXiv API, PubMed API, and Semantic Scholar API as real-time backends. All added 300ms-2s per query. For batch verification of 20 claims, that is 6-40 seconds of waiting. Pre-indexed FTS5 returns results in 50-200ms regardless of claim count.
The self-growing database was the most important architectural decision post-feedback. ASET never permanently says "no papers found" — it fetches, stores, and grows. The system gets smarter with every query.
Groq LLaMA 3.3 70B with key rotation handles rate limits gracefully — when one key hits its limit, the system automatically switches to the backup key without any user-facing error.
What we would do differently:
Build the multi-mode architecture from day one. The single-claim verifier was the right MVP, but document and YouTube modes are what make ASET genuinely useful for the people who need it most.
We also realized verification should happen where users already consume information — not only inside a standalone platform. This led to the development of the ASET browser extension, enabling real-time claim verification directly on webpages without breaking user workflow.


Built for the AWS 10,000 AIdeas Competition — Top 50 Finalist
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article