> sanjay.init()

Sanjay Raju

Data Scientist & AI Engineer

Data Scientist, AI Engineer, and Builder. Built ThisJustIn: a Gen-AI powered personalized news platform. Currently doing research on Speech2Speech models.

scroll

about

I'm Sanjay Raju - a Data Science graduate from Stevens Institute of Technology with a B.Tech in Artificial Intelligence from Amrita Vishwa Vidyapeetham. I build things at the intersection of AI, infrastructure, and product.

Most recently, I designed and built ThisJustIn from the ground up: a Gen-AI powered personalized news platform with E2E encrypted messaging, hybrid search, and a custom political alignment algorithm. The backend runs on AWS ECS/Fargate with a FastAPI service architecture, and the iOS app is built in SwiftUI.

Before that, I worked as an AI Engineer Intern at Tata Elxsi, building CUDA-accelerated C++ pipelines for video processing ML models. My research on lunar crater detection using deep learning was published in the Journal of the Indian Society of Remote Sensing.

Space exploration enthusiast
90 WPM typist
Tea over coffee, always
RDR2 & Ghost of Tsushima
Resume
Sanjay Raju

experience

Founder & Full-Stack Engineer

ThisJustIn · 2025 - Present

  • Designed and built a Gen-AI powered personalized news platform end-to-end
  • FastAPI backend with multi-role Docker containers on AWS ECS/Fargate, ALB, CloudFront CDN
  • Implemented E2E encrypted messaging using Signal Protocol patterns (X25519, AES-256-GCM)
  • Built hybrid search: PostgreSQL BM25 + pgvector HNSW + AWS Bedrock Titan embeddings + OpenSearch
  • Developed iOS app in SwiftUI/UIKit with Supabase RealTime for instant messaging
  • CI/CD pipelines with GitHub Actions and AWS ECR

AI Engineer Intern

Tata Elxsi · Jan 2023 - Jul 2023

  • Worked with a team of 30 on a multimillion-dollar video processing and interpretation project
  • Created a unified C++ pipeline for processing multiple ML models
  • Enhanced performance by implementing CUDA acceleration
  • Built multithreaded architecture for simultaneous input processing

Full Stack Developer

Al Qayam Electrical Equipments W.L.L. · 2024

  • Built website for a local business to market internationally in the Middle East
  • Designed pipelines to direct product enquiries and orders to company software

education

Master's in Data Science

Stevens Institute of Technology

Aug 2023 - May 2025

B.Tech in Artificial Intelligence

Amrita Vishwa Vidyapeetham

2019 - 2023 · GPA: 8.8/10.0

projects

Job Aggregator AI Agent

An intelligent system that parses resumes, scrapes 1000+ job listings, and matches them using AI embeddings with a conversational interface.

Python OpenAI Flask NLTK MongoDB
View Details

Lunar Crater Detection

Deep learning model detecting craters on the lunar surface from Chandrayaan-2 satellite imagery. Published in the Journal of the Indian Society of Remote Sensing.

Python Keras OpenCV YOLOv5 R-CNN
View Details

Image Captioning System

Multi-model system using LSTM and Feed Forward Neural Networks on the Flickr30k dataset, achieving 78% accuracy in image recognition and captioning.

Python LSTM Deep Learning Flickr30k
View Details

Mental Health in Tech Analysis

Deep analysis of survey data correlating workplace conditions and mental health in tech using statistical and Bayesian methods.

Python Statistics Bayesian Data Analysis

ML Video Processing Pipeline

Large-scale video processing pipeline using C++ for multiple ML models with CUDA acceleration, built with a team of 30 at Tata Elxsi.

C++ CUDA Computer Vision Multi-threading

Driver Drowsiness Detection

Real-time system using OpenCV for facial landmarks and a CNN model to detect drowsiness with early yawning warnings.

Python OpenCV CNN Real-time

blog

Building a Hybrid Search Engine with BM25, Vector Embeddings, and Real-Time SSE Streaming

Search in a video-first content platform is fundamentally different from traditional web search. Users type "Israel ceasefire" and expect not just matching videos, but an AI-generated news briefing, relevant creators, trending topics, and all of it appearing progressively - creators first (100ms), then videos in chunks, then an AI summary streaming token-by-token like a ChatGPT response. A single monolithic JSON response would mean 3-5 seconds of staring at a blank screen.

Read More

The 7,000-Line UICollectionViewCell: Building a TikTok-Style Video Feed That Doesn't Stutter

How I built a 4-tier video caching system and a single UICollectionViewCell that handles playback, engagement tracking, gesture recognition, trust visualization, recovery from compositor crashes, and first-frame detection for HLS streams - all at 60fps.

Read More

How I Built Signal-Level E2E Encrypted Messaging

Building a Signal-level encryption into a consumer app: A Technical deep dive

Read More

contact

terminal

$ echo $EMAIL

sanjjayrj@gmail.com

$ echo "Let's build something."

Let's build something.