Bridging the gap between theoretical AI research and production-grade software. I build systems that think, reason, and scale (usually without catching fire).
Technology isn't just about code; it's about mitigating chaos.
My journey began with a fascination for how data can shape the physical world. This drove me to C-DAC, where I didn't just build apps (I mostly fought fires). I helped modernize India's rural infrastructure monitoring, touching the lives of millions through the 'Quality First' & 'Meri Sadak' initiative. (Which, let's be honest, mostly involved explaining to stakeholders why "cloud" didn't mean actual clouds.)
Currently, I am diving deeper into the "brain" of modern computing at Arizona State University. My focus is on the subtle intricacies of Large Language Models.
Advanced coursework in Agentic AI, Natural Language Processing, Software Development, and Security. (Mastering the architectural constraints of scaling intelligence across decentralized nodes.)
Comprehensive curriculum covering the fundamentals of Computer Science and Engineering. (Building the foundational matrices for structural logic and gradient optimization.)
The problem? Ensuring data flowed reliably from remote villages where signal bars were hallucinations. By managing real-time GPS tracking and analytics for 98,500 km of rural roads, we had to cut latency without reliable networks. (No pressure, right? Just the connectivity of an entire subcontinent riding on my git commits.)
My insight? Don't trust the network. I directed full-stack engineering (Flutter, .NET Core, MySQL) with an aggressive offline-first REST API and JWT authentication. It allowed 150,000+ users to sync their work seamlessly, reducing field reporting latency by 40%.
I walked into a startup where the messaging infrastructure was a house of cards. I re-architected the mass-mailing platform (Ruby on Rails, AWS, Docker), completely automating HR workflows and cutting manual processing by 75%.
I also optimized the mobile media pipelines. By integrating FFmpeg video compression in Dart, I successfully reduced payload sizes by 50% and boosted low-bandwidth upload success rates by 40%.
Mapping a country manually is an impossible task. We had terabytes of satellite imagery but no way to process it fast enough.
I deployed a U-Net architecture on HPC GPUs, effectively teaching the computer to "see" roads with 96% accuracy. This accelerated urban planning workflows by 70%.
LLM agents often get stuck in loops or lose context during complex tasks. To fix this on Tau-Bench, I developed the PEVAL LangGraph multi-agent framework. (Because an LLM that can't remember what it did 5 minutes ago isn't an agent, it's a goldfish.)
I built a Memory Kernel and a Global Learning Node. Benchmarking the Qwen3 family via vLLM on A100 GPUs, the 32B model's consistency metric shot up from 0.0% to 30.0%.
GUI automation with Vision-Language Models often fails due to microscopic visual drift. You tell it to click a button, and it clicks empty space. (Which is fine until it accidentally clicks "Delete All" instead of "Save".)
I designed a Tri-Agent VLM pipeline on Intel Gaudi HPUs using Qwen3-VL, utilizing parallel coordinate clustering. This boosted OmniACT Action Scores to 51.5% and suppressed the AgentHARM Attack Success Rate to a mere 2.9%.
Orbit is a chaotic minefield, and standard algorithms are too slow to survive it. I trained PPO Agents (Reinforcement Learning) to autonomously dodge debris, treating orbital mechanics like a high-stakes chess match.
I also built OrbitGPT, a RAG-based 'Space Lawyer' that cites international treaties in real-time. (Basically, I built a simulation to break things safely, because debugging in orbit is... expensive.)
Standard LLMs obsess over nouns and starve the "connector" words (like "therefore"). It's like trying to build a bridge with only bricks and no mortar. (They sound confident, but they have the logical reasoning skills of a toddler.)
My fix? A surgical intervention on Llama 3.2. I identified 150+ discourse markers and amplified their gradients.
Using a massive 70B parameter model to judge a tiny 0.5B model is like using a sledgehammer to crack a walnut. It works, but it's wasteful. (And looking at the GPU bill made me physically ill.)
I pioneered a lightweight alignment pipeline for Qwen2.5 using a programmatic, heuristic reward function. 43.5% accuracy on GSM8K.
Mapping a country manually is an impossible task. We had terabytes of satellite imagery but no way to process it fast enough.
I deployed a U-Net architecture on HPC GPUs, effectively teaching the computer to "see" roads with 96% accuracy. This accelerated urban planning workflows by 70%.
Synthesizing latest research and technical insights from Medium.
Initiate a transmission. Let's build something extraordinary.