Mohnish Harwani

Multi-agent research assistant that accesses 1.6M+ data points and 340+ published software tools on nanoHUB, plus HPC clusters, to perform autonomous multi-domain research across biology, physics, and nanotechnology. Produces validated, reproducible results and scales with growing scientific contributions.

Active learning loop paired with public physics simulation data to accelerate materials discovery, achieving a 10x speedup in optimal alloy candidate identification. Published in Computational Materials Science, Elsevier.

Cost-efficient LLM pretraining using learned optimizers. Achieved a 30% performance gain versus Adam across several zero-shot tasks, improving zero-shot transfer and robustness across downstream tasks.

Meta-training pipeline integrating Sharpness-Aware Minimization into learned optimizer training, improving stability and zero-shot transfer across architectures, tasks, and initializations.

Intelligent health calendar prototype that generates personalized fitness and medication schedules using Kaiser clinical records and Apple HealthKit data. Implemented a privacy-first architecture using Apple Foundation Models with all health data processed and stored fully on-device.

Zeroth-order gradient estimation to accelerate training and improve generalization. CMA-ES-based warmup optimizer that outperforms SGD in early training for noisy, nonconvex loss landscapes.

Robust benchmarking suite to compare generalization across optimizers for neural networks. Designed evaluation protocols to measure optimizer stability, consistency, and rollout quality across diverse tasks and architectures.

Data structure combining Splay Trees with Hash Tables to accelerate access of nearby and recently queried information in O(log n). Exploits temporal and spatial locality to reduce amortized lookup cost.

In-memory hierarchical pseudo-filesystem supporting efficient retrieval using Heaps. Implements core filesystem operations (insert, search, delete) with optimized access patterns for hierarchical data.

Graph algorithms suite implementing SCC detection, maximum spanning trees, and constrained pathfinding for clustering and network optimization problems.

Stack-driven DFS and queue-based BFS traversal for generating and solving grid-based mazes. Implements both iterative deepening and breadth-first strategies for pathfinding.

Message board web application built for CS 180 at Purdue University.