Ravi Kumawat

I am an AI Engineer specializing in architecting, developing, and deploying scalable software systems that solve complex business problems and drive end-to-end automation. My core strength lies in bridging the gap between cutting-edge AI research and production-ready engineering — turning prototypes into stable, high-performing systems that real users depend on daily.

In my recent role as a Software Engineer and Founding Engineer at evolvue.ai (KrishAI Technologies Private Limited), developed under the guidance of Krish Naik (YouTube) and Mayank Agarwal (YouTube), I built and shipped several high-impact systems from the ground up — an AI-powered automated interview platform, an intelligent Udemy query response system, and Xenodocs (a document intelligence tool).

My technical focus revolves around applied AI — specifically bringing Agentic AI, Large Language Models (LLMs), and Retrieval-Augmented Generation (RAG) into stable, high-performing production environments. I am highly driven by building practical, automated workflows and am actively open to freelance projects and collaborations where I can translate cutting-edge computational techniques into tangible business value.

Prior to my work in the enterprise tech space, I served as a Research Intern at the Indian Institute of Technology (IIT) Ropar, where I led the initial build of IntelliClinix — a clinical-workflow tool that helps doctors at PGIMER Chandigarh run inference from trained DL models on histopathology and radiology images, and re-annotate poor outputs to drive model fine-tuning.

On the academic front, I am currently pursuing a Master of Data Science at the University of Melbourne, after completing my B.S. in Data Science and Applications from the Indian Institute of Technology (IIT) Madras.

At evolvue.ai, I was responsible for building the core technical infrastructure of an AI-powered automated interview platform from scratch — a system that conducts real-time mock interviews with candidates using voice-based AI. This wasn't a chatbot wrapper; it was a full production system handling concurrent interview sessions with audio streaming, intelligent evaluation, and detailed feedback generation.

Core technical work I delivered:

• —LLM Integration & Prompt Engineering: Designed and implemented the complete LLM orchestration layer — structured prompt pipelines for question generation, answer evaluation, and feedback synthesis. Built custom evaluation rubrics that produce consistent, actionable interview feedback.
• —Speech-to-Text & Text-to-Speech Pipelines: Engineered real-time audio processing pipelines — STT for transcribing candidate responses with high accuracy, TTS for generating a natural interviewer voice. Optimized latency to maintain conversational flow during live sessions.
• —AWS Infrastructure (S3, ECS, ECR): Architected the complete cloud deployment — containerized services on ECS with images managed through ECR, audio and document storage on S3, and RDS for persistent data. Designed for horizontal scalability and cost-efficiency.
• —CI/CD Pipeline: Built continuous integration and deployment pipelines using GitHub Actions — automated testing, container builds, and zero-downtime deployments to ECS. Every push to main triggers a complete build-test-deploy cycle.
• —Load Testing with Locust: Wrote comprehensive load testing suites simulating hundreds of concurrent interview sessions. Identified and resolved bottlenecks in the audio pipeline and LLM response chain before they hit production.
• —Monitoring with Sentry: Integrated Sentry for real-time error tracking, performance monitoring, and alerting. Set up custom dashboards to track LLM response times, STT accuracy rates, and session completion metrics across all services.

The platform was built under the guidance of Krish Naik and Mayank Agarwal, both recognized leaders in India's AI and data science community. Working closely with them gave me deep exposure to building AI products that scale — not just technically, but as businesses.

IntelliClinix is the project I built at IIT Ropar — a research-focused effort developed in close collaboration with clinicians at PGIMER, Chandigarh, one of India's largest referral hospitals with a daily footfall of more than ten thousand patients. Our lab had, over time, developed several deep learning solutions for medical images — histopathology whole-slide images (WSI), fluorescence in-situ hybridization (FISH), and MR/CT scans of organs — and we wanted to pilot these solutions inside the clinical workflows of the partners with whom they were originally developed.

I did the initial end-to-end build of the application myself, and I now guide a team of current students remotely who are continuing development. The resulting software is generic for both 2D and 3D medical images, lets clinicians run inference from trained models inside their normal workflow, and — crucially — lets them flag, correct, and re-annotate poor model outputs so the corrected data flows straight back into a model fine-tuning loop.

What I built and shipped:

• —Local-server desktop application: Designed and built the application as a web app that runs on a local server inside the hospital — important because patient data cannot leave the clinical environment. This required deep coordination between web-app engineering and on-premise deployment constraints.
• —2D & 3D image support: The viewer and inference pipeline are generic — the same app handles 2D histopathology WSIs and 3D MR/CT volumes, with consistent interactions for slicing, zooming, and overlaying model predictions.
• —MONAI inference engine: Integrated MONAI as the core medical imaging deep learning framework, wiring up lab-trained PyTorch models so clinicians can run inference on incoming studies with a single click.
• —nnU-Net pipeline integration: Plugged the nnU-Net training/inference framework into the workflow so segmentation models for organs and lesions can be retrained on corrected data without rebuilding infra from scratch.
• —CVAT annotation loop: Integrated CVAT as the annotation surface clinicians use to mark up poor predictions — those re-annotations feed directly into the fine-tuning dataset, closing the loop between deployment and model improvement.
• —3D Slicer & ITK-SNAP interop: Built interop with 3D Slicer and ITK-SNAP so radiologists who already live inside those tools can pull cases, overlay AI predictions, edit, and push corrections back without leaving their existing environment.
• —Clinician feedback → model fine-tuning: Designed the data flow so that every clinician correction becomes a labelled training example. This turns day-to-day clinical use into a continuous, in-situ training signal for the lab's models.
• —Remote mentorship & team handoff: After the initial pilot, I transitioned into a mentorship role — I now guide the current student team on architecture, model integration, and clinical UX while they push the project forward.

The work was the result of several months of iteration, feedback, and changes driven by the clinicians themselves. It required significant expertise in both web application development and understanding clinical workflows — the kind of dual-track engineering where a misplaced UI control or a slow inference call directly impacts whether a radiologist will actually use the tool. Once fully rolled out, IntelliClinix is expected to help deploy several of our lab's DL solutions at scale and benefit thousands of patients every month at PGIMER Chandigarh.