Aditya Dewan

Clubs: Quant Club (Goldman Sachs Quantathon: 100%), CMU Racing (built CUDA SVMs for 2x speedup). ACT 36.

Systems Work: Built Malloc (heap allocator), a Blackwell-optimized GEMM kernel, and a concurrent thread-safe file system from scratch.

• Designed a Poisson-process packet-loss probe scheduler with provable guarantees: modeled exponential probe inter-arrivals and cast repeated-anomaly detection as a geometric process, setting the probe rate for at least 95% detection within 7 recurrences; truncated the heavy exponential tail via a Chernoff bound solved with the Lambert W function, provably preserving 95% of the sampling mass while eliminating unbounded waits.
• Built a real-time C#/.NET diagnostics client (STA/MTA multithreading, async probes, rotating ndjson sessions) streaming 51 live client metrics; embedded statistical suite (Gaussian fitting, pairwise correlation heatmaps, scatter/histograms, session replay) saves roughly 598 cumulative hours of production troubleshooting time; integrated and shipped metric samplers into Amazon WorkSpaces (now in production) for clients including Ferrari, Intuit, Netflix, and Volkswagen.
• Architected an end-to-end anomaly detection and time-series forecasting ML pipeline: built synthetic data generation and validation infra injecting CPU/GPU/RTT anomalies, fine-tuned a Tiny Time Mixer with anomaly-detection and forecasting heads plus stream-pooling; added root-cause diagnostics and troubleshooting suggestions via a routing layer; entirely on-device inference and training with near-zero CPU client cost, augmented via hysteresis and debounce.

• Implemented a novel skip-connection Mamba model eliminating MHA-based retrieval degradation in SSMs; diagnosed Mamba-2 gather-aggregate bottlenecks on MMLU via embedding perturbation and eigenspace projection analysis, enabling efficient linear-time LLMs.

• 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.

• Devised a novel knowledge distillation algorithm at 16 achieving 94.8% accuracy on a 6,430-parameter malaria-classification CNN, beating standard distillation by 24.7 points (p = 6.3e-87); also delivered 2-24% accuracy gains in as little as 10% of training time across MNIST, CIFAR-100, and TinyImageNet.
• Reframed distillation from output-replication to using weak-teacher ensembles as curvature/domain proxies via Hessian eigenvalue spectral density, aggregated into an Approximated Saddle Region and injected into the student for re-exploration of low-loss saddle regions. PyTorch, PyHessian, NumPy. Awarded by NSA, ISEF, WAICY.
• Published as arXiv:2607.23346; ISEF project page.

• Architected the company's ML infrastructure (prompt-chunking, novel 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.