Aryan Jain

aryanj@andrew.cmu.edu/LinkedIn/GitHub

I'm a senior at Carnegie Mellon studying Computer Science and Machine Learning. I've worked on AI infrastructure at Tesla Autopilot/Optimus, Notion, and AWS, as well as applied ML at various startups.

Some of my hobbies include hiking, watching basketball (go Bucks!), reading, and most recently, playing pickleball. I also enjoy going on road trips, traveling abroad, and finding scenic spots.

Machine Learning

- Predicts post-RLVR lift before training using a LightGBM model trained on GRPO rollout features
- Reached 78% batch-ranking accuracy and 34% faster NLL convergence fine-tuning Qwen 3 on H100s
View project →
- CUDA-accelerated QNN modules and ops with full autograd support, written for a from-scratch deep learning framework
- Outperformed classical networks by 21–42% on 3 entanglement detection tasks by exploiting phase information
View project →
- LSTM network predicting Apple's closing stock prices, reaching a 4.93% MAE and an R² of 0.97
- Processed Yahoo Finance time-series data by normalizing features and building 100-day sliding windows for sequential input
View project →
- Multi-agent LLM system built on Google Gemini that does chain-of-thought reasoning with recursive subproblem validation
- Added agent backtracking, prompt decomposition, safety and correctness validation, and output critique
- Tested the chat interface on 20+ programming puzzles for a ~80% increase in passed tests over the base model
View project →
- GitHub Copilot-style browser extension for LaTeX code generation, autocompletion, and explanations with LLaMA 2
- Real-time (<450ms) Python backend serving context-aware completions over LaTeX code and docs indexed with LlamaIndex
View project →
- ResNet-9 computer vision model identifying 38 plant diseases with 93% accuracy
- Iterated through 3 architectures, sourced and augmented the training data, and tracked runs with Weights & Biases
- Deployed with AWS SageMaker and presented at Carnegie Mellon's AI4ALL research program
- CNN that identifies and bounds common road objects in a scene, aimed at autonomous driving
- Benchmarked 3 approaches — Perceptron, Sliding Windows, and YOLO — reaching 95% accuracy with VGG-16

Systems

- Low-latency distributed system for remote file operations using a server, caching proxy, and interposition library
- Cut file access latency 70% under simulated load with an LRU chunk cache hitting a 90% hit rate
- Server-proxy protocol over Java RMI handling 300+ concurrent clients with consistent open-close semantics
View project →
- Dynamic memory allocation system in C using segregated free lists, granular fit algorithms, and block-level optimizations
- Hit the 98th percentile for utilization and throughput — 22 million operations per second at 74.4% utilization
View project →
- Co-authored paper reviewing mathematical methods for extracting structure from common data representations
- Treats digital images as a dense, highly redundant medium and focuses on algorithms that exploit linear-algebraic structure for sizeable compression ratios
- Defines the mathematical tools behind each algorithm, then compares the techniques quantitatively and visually on a sampled data set
View project →
- Real-time pipeline for fire detection and location-based evacuation routing within 3 seconds of a fire outbreak
- Built a prototype ESP8266-based fire detector streaming sensor data over AWS IoT, with S3 and Lambda triggering alerts
View project →

Photo Blog

Madison, WIAugust 2026

Point LobosAugust 2026

See all 4 posts →