Aditya Dewan

Dean's List, High Honors (Fall 2025). ACT 36.

Clubs: Quant Club (Goldman Sachs Quantathon), CMU Racing (SVM midline generation rewritten on CUDA in PyTorch, 2x speedup).

Systems Work: malloc heap allocator and concurrent thread-safe file system in C; Blackwell-optimized CUDA GEMM kernel.

• Designed a Poisson-process packet-loss probe scheduler: treated each anomaly recurrence as a Bernoulli trial (recurrences to first detection are geometric) and set the probe rate for at least 95% detection within 7 recurrences; capped the inter-probe gap at a Chernoff cutoff (closed form via the Lambert W function) that clips at most 5% of gaps.
• Built a C#/.NET diagnostics client (STA/MTA multithreading, async probes, rotating ndjson sessions) streaming 51 client metrics, with correlation heatmaps, histograms, and session replay; estimated to save at least 598 hours/year of production troubleshooting (one hour per historical escalation ticket). Its metric samplers are in production in Amazon WorkSpaces for clients including Ferrari, Intuit, Netflix, and Volkswagen.
• Built an end-to-end anomaly detection and time-series forecasting pipeline: synthetic data generation and validation infra injecting CPU/GPU/RTT anomalies; fine-tuned a Tiny Time Mixer with anomaly-detection and forecasting heads and stream pooling; root-cause diagnostics via a routing layer; on-device inference and training with the model under 3 MB and under 2% client CPU including the app, with hysteresis and debounce.

• Implemented a skip-connection Mamba model for the drop in retrieval when multi-head attention is replaced by SSMs, toward linear-time LLMs.
• Diagnosed Mamba-2 gather-and-aggregate bottlenecks on MMLU with embedding perturbation and eigenspace projections; ran hybrid (Llamba) head-replacement experiments.

• Built a React web app simulating textual Markov Chain models for 900+ professionals worldwide; co-credited creator, cited in multiple ACM/CHI publications on AI literacy education (demo).
• Automated graph generation and n-gram estimation over Wikipedia-scale corpora via dynamic rendering and caching.

• Knowledge distillation algorithm started at 16. Reaches 94.8% accuracy on a 6,430-parameter malaria CNN distilled from a 70% weak teacher: 24.7 points over response KD and statistically equivalent to the from-scratch student (94.5%); up to 8 points over scratch controls on CIFAR-100 and MNIST at a matched epoch-10 checkpoint; 5 trials per experiment.
• Uses weak-teacher ensembles as curvature proxies via Hessian eigenvalue spectral density (top eigenvalues by power iteration), aggregated into an Approximated Saddle Region and injected into the student; also ResNet-101 to ResNet-18 on TinyImageNet. PyTorch, PyHessian. Awarded by NSA, ISEF, WAICY.
• Published as arXiv:2607.23346; ISEF project page.

• Built the company's first ML infrastructure (prompt chunking, drug-vector embeddings for medical-grade outputs); deployed an EC2 few-shot LLM drug-monograph summarization API and React app for in-clinic patient diagnosis. Python, Flask, ChromaDB, LangChain.

• Led a team of 3 shipping a recommendation-engine API to 45K clients computing principal levers of average-rating-change maximization for firm training programs; engineered a hybrid GAN / Gradient-Boosted Tree / MLP ensemble for production.