I'm a Computer Science student at Sardar Patel Institute of Technology, Mumbai, working across LLM inference efficiency, computer vision, and agentic systems. I care about the mechanics underneath the model — sparse sampling, attention geometry, agent orchestration, and where a system quietly breaks.
Published independent researcher (Ro-SVD, Zenodo 2026). Currently building Zenso, a video engagement analytics platform, and Sentinel, a multi-agent LLM red-teaming harness.
research(x) ≈ the same math that governs chips, choices, cities, and careers
— applied wherever it's actually useful.
|
Zenodo, May 2026 Online activation-level KV cache compression using position-aware truncated SVD on pre-RoPE key tensors in Grouped-Query Attention architectures.
|
|
Video engagement analytics platform. E(t) engagement scoring with hemodynamic lag correction, six-module LangGraph agent architecture, sparse-frame sampling (16 frames holds >0.99 Pearson r vs. full density at 7.5x speedup).
|
Multi-agent harness (LangGraph + OpenRouter/NVIDIA NIM) probing LLMs across five attack categories, with a cross-modal extension measuring how jailbreak success shifts across text, image, and video delivery — the "modality gap."
|
|
SkillFlow AI Agentic peer skill-bartering platform built for the JPMC Agastya Foundation. Autonomous matching between learners and mentors, no manual moderation.
|
CivicFlow Automated form discovery and filling agent using LangGraph, Playwright, and Claude Vision, with an RL feedback loop for adaptive form handling.
|
Automated Form-Filling Agent LLM agent that parses a target form's structure and generates Selenium automation to autofill it from a stored user profile. Built for FAR AWAY 2026.
|
|
Mineguard — YOLO Rockfall Detection Safety-critical rockfall detection on limited annotated datasets; precision-recall analysis across IoU thresholds.
|
Willovate Talent Hunt Production-grade ML service shipped under hackathon time constraints.
|
Transformer From Scratch Multi-head self-attention, causal masking, positional encoding, and autoregressive decoding implemented from first principles.
|
AI/ML Intern — Daten & Wissen Pvt. Ltd. · 2025 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.
- Hugging Face — Agents Course, Unit 1 Fundamentals Certificate
- Hugging Face — NLP Course Certificate
- Google Cloud Skills Boost — Agent Fundamentals
- Google Cloud Skills Boost — Use Machine Learning APIs on Google Cloud (skill badge)
- NVIDIA Deep Learning Institute — Agentic AI Explained
- Docker Essentials: Container Networking
Languages
AI / ML
Backend
Data
Cloud & Tools
Snake animation generated automatically from contribution history — see .github/workflows/snake.yml.
- Latest posts sync automatically via
.github/workflows/latest_articles.yml— run the workflow once after setup to populate this section.
- Publish Sentinel's modality-gap results as a public benchmark
- Ship Zenso beta to first external users
- Extend Ro-SVD to adaptive-rank compression
- Cross the 20-post mark on Gradient: Weight & Meaning
