I build production-minded AI systems — GPU compute infrastructure, evaluation and safety layers for LLMs, and developer tools for AI agents. Final-year CS student, graduating May 2027.

• Building Kestrel, a multi-tenant GPU compute platform on Kubernetes — Kueue admission with per-tenant quotas, pluggable placement policies (first-fit, bin-packing, priority) behind a reconcile loop, GPU-second metering, and a queue-depth autoscaler with scale-to-zero
• Built kea, a real-time incident root-cause engine — 🏆 1st place, OpenAI Codex Community Hackathon. A deterministic engine ranks root causes over a live service graph (Redpanda + Neo4j); the LLM only explains, with evidence-grounded checks and a benchmark against LLM-only baselines
• Creator of AntiVibe, a Claude Code skill that auto-generates educational deep-dives on AI-written code — 1,100+ GitHub stars, part of Mintlify's OSS Program
• Built SentinelLM, an open-source LLM proxy running a 7-stage concurrent safety evaluation pipeline (prompt injection, PII, toxicity, hallucination, relevance) across OpenAI, Anthropic, and Gemini
• GSoC 2026 with RUXAILAB (Uramaki Lab) — built the Focus Group Management and Analysis Module: real-time multi-participant sessions on LiveKit SFU and NLP-assisted thematic analysis (30 PRs, 24 merged)
• Product Developer Intern at Koncepts (PrepAiro), an AI EdTech platform serving 200K+ learners — shipped 165+ PRs to production, including AI-generated assessment builders and core backend systems
• Selected as 1 of 6 presenters at India Builds with Claude (Anthropic × Razorpay × Peak XV, Bengaluru) out of 2,300+ applicants
• Creator of Voidwalker, a Chrome extension + local MCP server exposing 16 custom tools with 256-bit token auth
• Founding Chair, IEEE Oceanic Engineering Society Student Chapter @ CVV (2024–2026)
• Contributed to OmegaUp, shipping the Profile READMEs feature end-to-end across the database, PHP backend, and Vue frontend

Currently working on LLM inference engineering, GPU compute infrastructure, and reward hacking in RL-trained models.