About

I studied Electronics and Communication Engineering at Amrita Vishwa Vidyapeetham in Coimbatore, where my final-year research on IoT-based crop monitoring was published at IEEE ICESC 2022. That project was my first real contact with systems that had to work reliably in the field — sensors failing, networks dropping, data needing to arrive on time.

After graduating, I joined Ericsson in Bangalore working on LTE/NR protocol stacks. Six months in, I noticed how much manual effort went into telecom testing and diagnostics. That observation stuck. I moved to the US for a master's in Software Engineering at RIT, where I built search platforms, studied ML, and started thinking seriously about how retrieval systems and language models could be applied to telecom infrastructure.

I joined Nokia's AI Platform Engineering team in Sunnyvale in August 2024. The interesting problems lived at the intersection of telecom systems and machine learning, and there weren't enough people who understood both sides. That gap became my focus.

At Nokia, I build production AI platforms that accelerate 5G product development. My primary project is an AI Test Automation platform that converts hundreds-of-pages telecom PRDs into executable regression plans using LLM-guided parsing and hybrid retrieval (FAISS + BM25 + reciprocal rank fusion). It's used by multiple internal engineering teams.

Outside of work, I created MCP-Telecom — the first Model Context Protocol server for network equipment. It lets AI agents safely interact with routers and switches across 7 vendor platforms via SSH, NETCONF, SNMP, and gNMI. It's published on PyPI.

I'm also researching proxy metric validation for A/B testing through PROXIMA, a framework that quantifies how much you should trust a given proxy metric, down to the segment level. The common thread across all of this: making AI systems that can be trusted in high-stakes infrastructure.

I prefer low-level understanding before abstraction. Before I use a library, I want to understand what it does at the layer below. Before I trust a metric, I want to see the query that produced it. Before I ship an AI feature, I want to know what happens when the model is wrong.

I'm skeptical of uncritical agreement and suspicious of systems that can't explain their own outputs. The safety gate in MCP-Telecom exists because I assume the LLM will confidently suggest something destructive — the question is how often, not if. PROXIMA exists because I saw teams trust proxy metrics that hadn't been validated below the aggregate level.