I'm a Computer Science student at Sardar Patel Institute of Technology, Mumbai, working on
LLM inference efficiency, computer vision, and agentic systems. Most of what interests me lives
one layer below the API — sparse sampling strategies, attention geometry, agent orchestration,
and the exact place a system quietly breaks under load.
Published independent researcher (Ro-SVD, Zenodo 2026) and former AI/ML intern
at Daten & Wissen. Currently building Zenso, a video engagement analytics
platform, and Sentinel, a multi-agent harness for red-teaming LLMs across text,
image, and video delivery.
Four lines of work, roughly in the order they started.
Position-aware truncated SVD on pre-RoPE key tensors for Grouped-Query Attention. Evaluated on 1,602 LongBench samples with Qwen2.5-7B-Instruct: 73.91% ROUGE-L win rate vs. uncompressed baseline, 35.1% KV-cache memory reduction. Custom CUDA de-rotation kernel with cuBLAS batched low-rank projection.
End-to-end pipeline for inverse RoPE de-rotation, truncated SVD, and progressive online re-compression. 7-configuration sweep over 1,602 LongBench samples tracking ROUGE-L, perplexity, latency, and GPU memory. Explored Triton-based batched randomised SVD to cut the 29.5% Python-loop overhead.
E(t) engagement scoring with hemodynamic lag correction. Six-module multi-agent architecture on LangGraph. In active development on private infrastructure (Kaggle/local), pending IP protection.
Multi-agent harness (LangGraph + OpenRouter/NVIDIA NIM) probing LLMs across five attack categories, extended to measure how jailbreak success shifts across text, image, and video delivery — the "modality gap." In active development on private infrastructure, pending IP protection.
Agentic peer skill-bartering platform built for the JPMC Agastya Foundation. Autonomous matching between learners and mentors, no manual moderation.
Neural engagement analytics platform built on sparse-frame video understanding and a six-module LangGraph agent architecture.
Multi-agent red-teaming harness for LLM safety, with cross-modal jailbreak measurement across five attack categories.
Automated form discovery and filling agent using Playwright and Claude Vision, with an RL feedback loop for adaptive form handling.
LLM agent that parses a target form's structure and generates Selenium automation to autofill it from a stored profile. Built for FAR AWAY 2026.
YOLO-based rockfall detection for safety-critical monitoring on limited annotated datasets.
Production-grade ML service shipped under hackathon time constraints.
Multi-head self-attention, causal masking, positional encoding, and autoregressive decoding implemented from first principles.
Fine-tuned PaddleOCR models for license plate detection and recognition (~88% end-to-end accuracy). Evaluated U-Net/SAM/ResNet segmentation backbones for production latency–accuracy trade-offs. Ran Optuna hyperparameter sweeps.