Tejas Saraf
I started my career solving real-time infrastructure problems at scale. At Jio Platforms, I designed a WebRTC and Agora SDK streaming pipeline serving 10,000+ concurrent attendees, built a Media SDK on MediaSoup supporting SFU, P2P, and MCU architectures, and reduced call drop rates by 30% contributing to $1M in annual product revenue. At Elite Softwares, I led the migration of a legacy monolithic CRM to microservices on Kubernetes and AWS EKS, cut system latency from 600ms to 80ms with Redis caching, and reduced cloud costs by 62% across three AWS regions.
More recently, I've been working at the intersection of AI and real-time systems voice agents, shipper-carrier calls end-to-end at sub-500ms latency and predicts churn risk 7× earlier. instaMIND detects security incidents in under 100ms entirely on-device. OpusAI compresses diabetic retinopathy screening from weeks to under 8 seconds.
I hold a filed patent in quantum hardware design and quantum algorithms for machine learning, and I've published research on quantum cryptography for UAV communication and AI-driven cloud application performance.
I'm looking for software engineer, ML engineer, or founding engineer roles where I can own real-time AI systems end-to-end. If that sounds like your team, let's talk: saraftejas12@gmail.com
Professional Summary
Software engineer with 3+ years of production experience building real-time streaming infrastructure, voice AI agents, and cloud-native backend systems. Architected WebRTC pipelines serving 10,000+ concurrent users at Jio Platforms ($1M annual revenue), led monolith-to-microservices migrations cutting latency from 600ms to 80ms, and deployed LLM-powered systems with 90% cost reductions. Built FreightVoice and instaMIND (on-device ML at sub-100ms latency). Patent holder in quantum algorithms for ML. M.S. in IT Management from Illinois Institute of Technology (3.9 GPA).
Employment History
The team's NLP pipeline was expensive (GPT-3 inference costs), slow (high response latency on RAG queries), and lacked scalable preprocessing for meeting transcripts at volume.
- Model Deployment: Trained and deployed a RoBERTa model on AWS SageMaker achieving 1.2x inference improvement, implemented unsupervised topic clustering on meeting transcripts, and integrated Spark for large-scale preprocessing and inference.
- RAG System: Deployed an AWS Kendra RAG system handling 2,000+ user interactions, reducing response time by 35% through semantic chunking, vector retrieval, and guardrails.
- Cost Optimization: Transitioned sentiment analysis from GPT-3 to Llama 3.1 via shadow mode A/B testing, achieving 90% cost reduction while maintaining an F1 score of 0.92.
JioEvents needed to scale from hundreds to 10,000+ concurrent attendees without compromising audio/video quality or adding latency — all while contributing to a product generating $1M in annual revenue.
- Streaming Pipeline: Deployed a real-time audio/video streaming pipeline by building a WebRTC and Agora SDK layer with RTMP fallback and a custom sliding window buffer system, serving 10,000+ concurrent attendees and directly contributing to $1M in annual product revenue.
- Media SDK: Developed a WebRTC-based Media SDK using MediaSoup, enabling seamless audio/video communication via SFU, P2P, and MCU architectures and reducing call drop rates by 30% across all event types.
- Authentication Reliability: Reduced OTP authentication error rate from 8% to under 1% within 6 weeks by building 200+ integration tests, implementing Test-Driven Development with 90% code coverage, and adding DataDog logging and alerting for early issue detection.
- CI/CD & Microservices: Designed Node.js microservices on Google Cloud Run, reducing multi-repo CI failures from 2,500/week to 20/week within one quarter through automated testing and CI/CD integration via GitHub Actions and Docker.
The company's CRM platform was a monolithic system struggling with 600ms response times, frequent security vulnerabilities, and cloud bills growing faster than revenue. It needed a ground-up infrastructure overhaul.
- Microservices Migration: Led migration from a monolithic CRM to a microservices-based architecture using Java, Spring Boot, React, Kubernetes (AWS EKS), and AWS App Runner, reducing security vulnerabilities by 50% and doubling service scalability with independent fault isolation.
- Latency Optimization: Improved system latency from 600ms to 80ms by implementing Redis for caching and asynchronous processing, achieving a 10x reduction in database load.
- Cloud Cost Reduction: Optimized AWS infrastructure across 3 regions, reducing cloud compute costs by 62% through rightsizing services and automated cost monitoring with AWS Cost Explorer.
- DevOps & Observability: Created Kubernetes deployments using Helm charts, Argo CD, and GitHub Actions for fully automated CI/CD pipelines; built production monitoring dashboards with AWS CloudWatch and Grafana for actionable operational insights.
Education
Coursework: Applied Artificial Intelligence and Deep Learning, Object-Oriented Application Development, Advanced Software Programming
Activities: YC Voice Agents Hackathon Finalist (FreightVoice) · Google DeepMind × InstaLILY AI Hackathon (instaMIND) · ScarletHacks 2026 Winner (OpusAI) · Vivpro.ai Hackathon Runner-Up (ClinSearch)
Coursework: Natural Language Processing, Data Structures and Algorithms, Object-Oriented Programming, Computer Networking
Activities: Patent Filed: Quantum Hardware Design and Fabrication and Quantum Algorithms for Machine Learning (202321086733)