MS Information Systems student at Northeastern University (Boston). I build across embedded systems, iOS, ML, and cloud infrastructure.
Portfolio: akshaymrdev.vercel.app
- EdgeSplit: Disaggregated LLM inference that prefills on a laptop CUDA GPU and decodes on an Android phone, handing off raw sequence state over a checksummed TCP frame. V2 Cuts time-to-first-token 65 to 82% on Qwen3-0.6B, and a hand-tuned NEON accumulator in llama.cpp's
Q4_K×Q8_Kdot product adds 4.8% decode throughput on the phone. Code · Demo - MBTA Live: Live MBTA arrivals, home-screen widgets, and lock-screen Live Activities on iOS, backed by a nine-service event-driven platform that stands up on EKS or GKE from one Terraform codebase. Code · Demo
- WatchGPT: Photo analysis and streamed chat on Apple Watch, behind a Go backend with Cognito auth, Redis caching, SSE streaming, and circuit breaking on EKS. Code · Demo
- Orbital Threat Assessment: ML pipeline for hazardous-asteroid detection over 958K NASA JPL records, reaching 0.99 PR-AUC on a 449:1 imbalanced held-out set. Code · Notebook
microsoft/typescript-go: Fixed a false TS1293 on destructured
requireunder--module preserveby exempting binding elements from the ESM syntax check. PR #4800google/tunix: Fixed a TOCTOU race condition. PR #1627
pytorch/pytorch: Fixed stale 404 documentation links flagged by the nightly URL lint job. PR #189619
ARM-software/ComputeLibrary: Replaced
std::dequewithstd::vectorin the HeuristicTree BFS to clear a GCC 14-Wstrict-overflowbuild failure. PR #1304
- A system and method for optimizing Bill of Material cost and power performance for Drones: co-inventor, IN Patent 582,124 (2026)
- A system and method for optimizing Bill of Material cost and power performance of Graphics Processing Unit (GPU) core: co-inventor, IN Patent 537,663 (2024) · Certificate
