Building systems, GPU code, and software I want to use. My projects live here, on a server I run.
Hey, I'm Aryan.
I build GPU kernels, systems that stay up, and software I want to use. My work spans systems programming, infrastructure, and AI — from C and CUDA to self-hosted services and local LLMs.
This is my own Git server, where I keep building and looking after the code myself.
Portfolio · LinkedIn · Public activity and contribution graph · All public repositories
Selected projects
-
Freeway — LLM Gateway — a self-hosted OpenAI-compatible gateway in TypeScript and Node.js. Routes across 12 providers with quota-aware failover, context refitting, exact-request caching, and optional semantic caching. Providers are configured through JSON, with no external runtime dependencies.
-
Vega — Multi-Agent AI Trading System — four research agents sharing one local LLM through MLX, with tool use, research budgets, and checkpointed outputs. A review agent can approve, reduce, or veto ideas; deterministic risk checks gate Alpaca paper-trading orders. Includes a dashboard for agents, research, positions, and execution.
-
GPT-2 Inference Engine from Scratch — a four-person CUDA C++ project implementing and profiling GPU inference kernels on an NVIDIA A40. Worked on KV caching, FlashAttention-2, shared-memory bank conflicts, and tile tuning using Nsight and CUTLASS. The team's optimized stack reached 24.2× end-to-end speedup. Project write-up.
-
Homelab — Self-Hosted Infrastructure — the infrastructure behind the services I run: Proxmox, a three-node high-availability K3s cluster, OPNsense, and Cloudflare, with Ansible automation, monitoring, and centralized logs. Atlas maps the architecture; its deployment uses Gitea Actions, container Packages, and Komodo on Orion. Explore Atlas.
-
TLS Edge Failover — an Oracle HAProxy relay and Python watchdog for my homelab. Checks primary, local, and backup HTTPS routes before switching DNS, then automatically fails back after sustained recovery.
-
Systems Programming — C and Linux/UNIX — coursework covering a memory allocator, an epoll-based TCP server, a POSIX shell, a parallel build tool, and local MapReduce. My allocator placed first out of 400 students. These course projects are described on my resume; their source is not linked here.
More to explore
- Distributed Systems coursework — group projects for CS 425.
- EnvSync — a CLI and web dashboard project for encrypted environment-variable collaboration.
- XRayPDF — a PDF inspection tool that reveals hard-to-see content with rendering, contrast enhancement, and OCR, and can export a flattened copy.
Why my code lives here
After my GitHub account was suspended, I recovered projects from local copies and backups and moved my work to this Gitea instance on my own server. Some older history and assets were missing from those backups, so the archive is still a little uneven. I'm continuing the work here.
Public activity
My live contribution graph and public event feed are on Gitea's Public Activity tab. The Overview tab shows this README.