6 builds · 2023 / 2026
Six builds across AI inference, compute optimization, 6G networking, distributed systems, and space. Each one starts from a real bottleneck and gets measured against a real baseline.
Built by Aryan Putta, Computer Science and Data Science at Rutgers University, class of 2028. 31 merged open-source pull requests across NVIDIA, IBM, Kubernetes, Pulumi, Microsoft, DeepSpeed, AWS Labs, LinkedIn, Hugging Face and simdutf. Open to off-cycle SWE internships, co-ops, AI infrastructure, ML systems, cloud/backend systems, and Member of Technical Staff internships. Reach me at aryan.putta@rutgers.edu.
A compute runtime that measures each workload at run time, then routes it to the right backend: scalar CPU, threaded SIMD, or CUDA. I ran it on real SuiteSparse matrices and SNAP graphs, and checked correctness against reference results before making any speed claim.
Serve an LLM across many GPUs and the workers keep redoing the same prefill work, wasting memory. NimbusMesh-X is a shared KV-cache layer that lets nodes reuse each other's cache, so time-to-first-token holds up when memory gets tight.
When long context has to fit a hard memory budget, most KV caches just drop old tokens and lose the evidence. EigenKache compresses the cold middle into attention-conditioned landmarks instead, so the model keeps long-range recall. I benchmarked it against window and H2O-style eviction under the same budget.
5G and 6G networks split limited spectrum across users with very different needs. I framed that allocation as a QUBO and solved it with quantum-inspired local search, against greedy and linear-programming baselines, under a shared latency model that includes radio, transport, queueing, and inference.
EEG screening models get scored offline, but real deployment means streaming two-second windows at 256Hz under a sub-100ms budget. This benchmarks accuracy and per-window latency together across architectures on the OpenNeuro ds004504 dataset, 88 subjects, leave-one-subject-out.
TCP assumes a link that behaves consistently. Space links don't. This models a non-ergodic channel with Gilbert-Elliott bursts, runs TCP over it, and detects when the link's behavior shifts so the transport can react instead of stalling.
