About
By day, I build ml infra. By night, I never quite know how I'll end up spending my time. It might start with an ML paper and end with me reading about intelligence, political institutions, or why a team made a questionable strategy call in F1. Occasionally, I'll find myself three tabs deep into astrophysics with no recollection of how I got there.
I like building things, breaking my understanding of them, and occasionally realising I had misunderstood the whole thing. There's something oddly satisfying about finally understanding a concept that seemed impossible a few days ago. There's also something deeply annoying about discovering that the explanation was much simpler than I made it out to be.
Outside of code, I'm interested in how minds work, how societies organise themselves, how economies behave, and how a few milliseconds can decide the outcome of a race. I don't necessarily have a grand theory connecting all of these things. For now, they're just things I like thinking about, probably more than is strictly necessary.
I'm still figuring a lot of things out. The reading list grows faster than it shrinks, half the ideas in my notes will probably never see the light of day, and I occasionally spend more time thinking about doing something than actually doing it.
Anyway, back to whatever I was supposed to be doing.
Work Experience
Skills
Recent work and projects
A collection of ML systems, research prototypes, and random experiments. Some are polished. Most aren't.
A minimal vLLM-style LLM inference engine built from scratch in pure PyTorch for a single GPU. Implements PagedAttention-style KV-cache management and continuous batching, using Qwen2.5-0.5B-Instruct as the reference model. Block manager, KV cache, sequence handling and paged attention are implemented and tested; scheduler, engine, server and benchmarks are still in progress.
A deterministic intent routing and execution framework for machine-to-machine commerce, built on Razorpay test-mode APIs. An untrusted LlamaIndex multi-agent layer handles discovery and cart building, while a rule-based policy engine (spend limits, idempotency, category rules, human approval) gates every payment action. Orders are only marked paid after a signature-verified Razorpay webhook, and every security-relevant event is written to an HMAC-signed, hash-chained audit trail.
A comprehensive multimodal pipeline for translating sign language videos into natural language text, implementing the mathematical architecture specified in the Comprehensive Architectural Specification: Mathematical Pipeline for Sign Language Translation Systems document.
A production-grade MLOps pipeline that predicts supply chain disruptions and recommends real-time mitigation strategies. This project combines prescriptive analytics with machine learning to help businesses minimize delays and optimize resilience.
Archived Failures.
I've participated in 5+ hackathons throughout college. There's something fun about turning an idea into a working product in 48–72 hours with a bunch of motivated people. Nothing beats the chaos of going from a random idea to a demo-ready project in a weekend. here are some of the hackathons I've participated in, and the projects I've built at them.
HackIndia Spark 9
Roorkee, Uttarakhand
developed a RAG system that simplifies access to legal information by helping users understand Indian legal documents, procedures, and regulations through natural language interactions. Leveraging vector-based document retrieval and semantic search, the system delivers context-aware responses grounded in relevant legal sources. With multilingual support and an intuitive web interface, the solution improves legal accessibility, enabling users to navigate complex legal matters more effectively and make informed decisions.
Shivalik National Level Hackathon
Dehradun, Uttarakhand
developed an AI-Powered Personal Health Companion that helps users proactively monitor and manage their health through AI-driven symptom analysis, biometric data interpretation, and personalized health insights. By combining natural language symptom understanding with health records and real-time access to medical professionals, the platform enables early risk detection, informed decision-making, and improved access to quality healthcare. The solution is particularly impactful for remote and underserved communities, where timely medical guidance is often limited. and secured 7th place.
Volkswagen i.mobilothon 5.0
Online
proposed an V2X based road safety system that detects potholes, open manholes, and other road hazards in real time using onboard cameras and fine-tuned YOLOv8 models. Detected hazards are automatically geotagged and broadcast through V2V/V2X communication using lightweight wireless protocols such as NRF or DSRC, enabling nearby vehicles to receive instant warnings even without internet connectivity. The system can integrate with existing infotainment systems or mobile applications to provide proactive driver alerts, reducing accident risks and improving road safety. Designed to be lightweight, scalable, and easily retrofitted to existing vehicles, the solution enables large-scale deployment with minimal infrastructure changes.
ML-alrIEEEna Hackathon
Graphic Era University, Dehradun, Uttarakhand
Participated in the ML-alrIEEEna Hackathon and cleared the first round, while working on predictive analytics on given problems but due to some unforeseen circumstances, I couldn't attend the second round of the hackathon.
Qhackathon 2026
Quantum University, Roorkee, Uttarakhand
developed ML Productionizer, a VS Code extension that automates the transition from experimental Machine Learning code to production-ready projects. Leveraging Large Language Models (LLMs) such as Gemini, Hugging Face Router, Ollama, and other local or remote backends, the extension analyzes single-file ML training scripts and automatically generates a scalable project structure following industry-standard software engineering and MLOps best practices. This enables developers to reduce manual setup, improve code maintainability, and accelerate the deployment of machine learning solutions.
Get in Touch
If You'd Like to Talk. Just send me an email at gourav.sahu.1695@gmail.com and I'll get back when I can. GitHub is linked below.