Aditya Dewan

Clubs: Quant Club (Goldman Sachs Quantathon: 100%), CMU Racing (built CUDA SVMs with 2x speedup for GPU midline generation).

Systems Work: Built Malloc (dynamic memory allocator) and a fully concurrent, thread-safe file system in C from scratch.

• Designed a Poisson-process packet-loss probe scheduler with provable guarantees, modeling inter-arrivals and anomaly recurrences to achieve at least 95% detection within seven recurrences; bounded heavy-tail wait behavior via Chernoff-bound analysis and Lambert W-based calibration.
• Built a real-time C#/.NET diagnostics client (STA/MTA multithreading, async probes, rotating ndjson sessions) streaming 51 live metrics with embedded Gaussian fitting, pairwise correlation heatmaps, scatter/histograms, and session replay; reduced production troubleshooting time by roughly 50% for clients including Netflix, the U.S. Army, and Volkswagen.
• Architected an end-to-end time-series ML pipeline that injects CPU/GPU/RTT faults, dynamically labels causal signals, and fine-tunes a Tiny Time Mixer with custom anomaly-detection and forecasting heads; implemented root-cause routing for targeted remediation.

• Implemented a novel skip-connection-based Mamba model to mitigate multi-head-attention-related retrieval decline in SSM architectures for linear-time LLM settings.
• Analyzed State Space Models and Mamba-2 gather-and-aggregate bottlenecks on MMLU retrieval benchmarks, including embedding perturbation/projection behavior.

• Engineered and deployed a React web application for simulating and visualizing textual Markov Chain models for 900+ professionals; used in a Concord University study (demo).
• Built efficient graph generation and probabilistic next-word computation pipelines for bigram, trigram, and tetragram models with dynamic rendering/caching for low latency on large corpora.

• Devised a loss-landscape distillation compression method (SPRKD) yielding 2-24% higher accuracy at roughly 10% of conventional training time in tested configurations.
• Computed high-descent-potential saddle points in the teacher optimization landscape via Hessian approximation and spectral analysis for rapid curvature-tracking student convergence.
• Project: SPRKD

• Engineered the company’s first ML backend infrastructure for high server volume and minimum latency using prompt chunking, custom drug vector embeddings, and safety guardrails.
• Developed an LLM-based drug monograph summarization API and web application for real-time clinical patient diagnosis workflows.

• Led a team of three to architect and deploy a recommendation engine API serving 45,000+ clients in prototyping, targeting maximal average rating-change uplift.
• Built a hybrid GAN/Gradient Boosted Tree/MLP model pipeline and production deployment stack with Python, PyTorch, TensorFlow, XGBoost, SQL, and Flask.