Education
New York University
Graduate coursework across ML systems and core CS: Operating Systems, Fundamental Algorithms, Deep Learning, Reinforcement Learning & Decision Making, Large Language & Vision Models, Big Data.
University of Mumbai
Foundation in ML, AI, NLP, computer vision, data structures & algorithms, and databases, with honors in Computational Finance.
Experience
Research Assistant — AI4CE Lab, NYU
- Working on the Wanderland project — large-scale, real-world data collection for scene-scale world models and 3D scene understanding.
- Building data-quality and processing pipelines to improve the training data underlying the lab's world-modeling research.
Research Assistant — CILVR Lab, NYU ↗
- Built data pipelines and AI workflows ingesting and transforming 5TB+ of data across 100K+ runs. (ICLR 2026)
- Built evaluation & observability pipelines deploying GPT-4o across 100K+ runs with reference-based metrics and structured validation — r ≈ 0.78 with ground truth.
- Wrote clean, well-tested Python integrating multiple models into pipelines, with code review and validation to catch issues early.
CS Tutor & Grader — NYU Courant
- Computer Science Tutor — CSCI-UA 421 Numerical Computing (Jan 2026 — May 2026)
- Computer Science Grader — CSCI-UA 201 Computer Systems Organization (Sep 2025 — Dec 2025)
Software Engineering Intern — NTT DATA Business Solutions
- Built production data pipelines with LangChain + GPT-3.5, integrating embedding-based retrieval via ChromaDB and REST APIs on AWS — improving retrieval accuracy by 40%.
- Optimized API consumption and latency to ~1.8s through token-efficient retrieval and top-k strategies.
Research & Publications
World-Model-Based Evaluation of Robotic Policies
Large-scale evaluation of robotic policies inside a video world model — 5TB+ of Open-X Embodiment data, 100K+ rollouts, and a GPT-4o-based multimodal judging harness reaching r ≈ 0.78 agreement with ground truth.
Tulsi Leaf Disease Detection Using CNNs
End-to-end computer vision pipeline for multi-class plant disease detection using transfer learning on CNN architectures (VGG16, ResNet50) over a custom-curated image dataset, with augmentation and HSV-based preprocessing for robustness.
High-Frequency Trading Using Sentiment Analysis
HFT strategy combining technical indicators (MACD, RSI, Bollinger Bands) with Twitter/Google News sentiment and an LSTM price forecaster — improving predictive accuracy by 14%, backtested via Sharpe ratio, drawdown, and returns.
Projects
Foundation Model Distillation ↗
DINOv2 teacher–student distillation with VICReg-style invariance/variance/covariance regularization and multicrop training — +5.1% absolute ImageNet k-NN accuracy; benchmarked SimCLR, BYOL, SwAV.
CogniLLM — ARC-AGI Reasoning ↗
Reasoning-augmented system with LoRA adapters, synthetic data, dual transduction/induction prompting, and a four-solver voting mechanism — task accuracy 8% → 29%.
World Model for Robotic Exploration ↗
Online latent-space world model blending curiosity-driven exploration with optimal trajectories; extended DreamerV3 with 32-env mixed action sampling and a masking fix for faster convergence under 10⁵-step budgets.
Agentic Web-Navigation AI ↗
Autonomous agent orchestrating LLM reasoning with browser tool use via multi-step planning/execution pipelines and trajectory logging — >70% task success over 100+ runs with guardrails and validation.
AI Meeting Copilot ↗
Real-time meeting assistant with ~300ms transcription latency, a two-window episodic memory architecture (90s verbatim + compressed history), and contextual suggestions in <2s. Live app →
Distributed Data Pipeline ↗
Fault-tolerant, horizontally scalable ETL pipeline ingesting real-time streams with partitioned event streams, stateful transformations, and data-quality checks at pipeline boundaries.
LLM-Driven SQL RAG Pipeline ↗
Natural-language-to-SQL workflow composing retrieval, prompting, and structured outputs over a relational banking database — +35% accuracy, −40% latency via evaluation-driven iteration.
TinyURL System Design ↗
Backend with 8+ REST endpoints, JWT auth, Base62 encoding (~56B key space), integration & performance tests, and CI/CD — indexed for low-latency queries under concurrent load.
AI Document Extraction Pipeline ↗
Production-style document-understanding pipeline for invoices, legal docs, and receipts — CUDA-accelerated inference cutting latency 30% across 50K+ documents.
ConstituAI — Llama-2 Fine-Tuning ↗
Fine-tuned Llama-2-7B-Chat on an Indian law dataset with LoRA/QLoRA and 4-bit quantization — +20% query accuracy, −70% VRAM on limited hardware.
Credit Risk Modeling ↗
Probability-of-default models over 1M+ financial records (44+ variables), with feature engineering, calibration, and precision-recall/risk-sensitivity evaluation — first-prize winner.
Fraud Detection & Cost Optimization ↗
Fraud models on 2M+ transactions under heavy class imbalance, plus a cost-sensitive threshold-tuning framework balancing fraud loss against false-positive user impact.
Crowd Dynamics Risk Prediction
Crowd-risk pipeline combining density estimation with optical flow to model dense crowd motion as fluid dynamics; fine-tuned RAFT/GMFlow on TUBCrowdFlow over 2M+ NWPU-Crowd annotations.
OS Resource Simulators (C++)
Discrete-event CPU scheduler, multithreaded I/O scheduler, and virtual-memory pager — FCFS/SRTF/RR, SSTF/LOOK/C-SCAN, Clock/NRU/Working-Set.
Trade Order Management API ↗
RESTful trade-order backend with Pydantic validation (+35% API response time), Dockerized on AWS EC2 with a full CI/CD pipeline and automated unit/integration tests.
Patient Readmission Prediction ↗
End-to-end ETL + ML pipeline on the 100K-record UCI Diabetes dataset — SMOTE for imbalance, XGBoost/LR/MLP models reaching 0.81 ROC-AUC, deployed via Streamlit.
Full-Stack Blogging Platform ↗
MERN blogging platform with JWT authentication, route protection, CRUD post management, and real-time commenting.