Mahyar Riazati.

hello

I work on web applications, developer tools, and reinforcement-learning research.

I like understanding how things work, then building something useful with that understanding.

What interests me most is the gap between something that works in an experiment and something that holds up in the real world—and the engineering it takes to bridge that gap.

I love computers ツ

That was the old site.
There’s a little more here now.

Fourier study / 𝓜z ∈ ℂ
z(t) = Σ cₖ e2πikt512 samples

A mathematical script M reconstructed from a complex Fourier series. Each contour is sampled at 512 equal arc-length intervals; 128 coefficients determine rotating circles and vectors that trace its outline. Animation pauses offscreen and is replaced with a static glyph when reduced motion is requested.

Selected work

open source

Tools I build and maintain. Source and implementation notes included.

XRAT

Networking / systems

A Swiss Army knife for proxy connections.

XRAT brings proxy configuration, connection testing, and runtime management into one toolkit. It helps find usable connections and keep track of how they perform.

Implementation notes

I develop the Rust CLI, terminal interface, web interface, and SDK. The toolkit parses subscriptions, tests HTTP latency concurrently, stores measurement history, and supervises proxy processes with health checks and automatic rotation.

Rust · CLI / TUI / Web · Xray-core · sing-box

OGym

Reinforcement learning / tooling

Reinforcement-learning environments, but in Rust.

OGym provides environments for reinforcement-learning experiments, from classic control tasks to physics simulations. A shared interface and typed action and observation spaces make them usable across different implementations.

Implementation notes

I develop the environment interfaces, batch execution, and implementations using native Rust, Rapier, and MuJoCo. The project also includes benchmarks against Gymnasium, with the methodology and raw measurements available alongside the results.

Rust · Typed spaces · Rapier · MuJoCo

Mahfel

Real-time / collaboration

A shared room, on your own infrastructure.

A self-hosted meeting and collaboration platform with a Farsi-first interface. An ongoing engineering project, with audio/video, shared notes, whiteboards, and recordings.

Implementation notes

I develop Rust services and a Next.js interface, connecting WebRTC media, application signaling, background jobs, and shared storage. The architecture separates API, media, and worker responsibilities.

Rust · Next.js · WebRTC · PostgreSQL

Research

reinforcement learning

Reward shaping, experience transfer, and when outside guidance helps.

SAFER

thesis

What if the extra reward is pointing the wrong way?

My thesis at the University of Tehran developed a method for incorporating uncertain guidance into reinforcement learning while preserving the original task objective.

The method works with both on-policy and off-policy algorithms:

  • On-policy: SARSA and A2C
  • Off-policy: Q-learning and discrete SAC

I implemented and evaluated the method in Rust, with Python tools for analysing experiments with helpful and adversarial guidance.

MSc research · 2022–2025

Experiments & collaborations

At Sharif’s Artificial Creatures Lab, I worked on experience transfer and reward guidance, contributing to Ex-RL and Deep ExRL.

At Tehran’s Cognitive Systems Lab, I explored LLM-generated rewards and how LLM guidance can help RL agents solve tasks. Along the way, I independently built an intelligent tutoring system (ITS) that later became RastAI .

Ex-RL experiment pipeline

Publicationscoauthored research

Experience

engineering / research / teaching

Selected context. The CV is the complete formal record.

RastAI

Jul 2025–present

Software engineering · Developer tooling · R&D

I played a central role in the MVP and new platform core, across application development, infrastructure, and deployment. My work includes Kotlin/Spring Boot, React/Next.js, and AI-assisted authoring and grading with scoped tools and human review.

I also build developer tools that make it easier for teammates to work with coding agents. These include Rastploy, a Go deployment CLI and daemon with scoped access, deployment logs, and rollback, alongside tools for runner maintenance, service checks, and configuration reconciliation.

Cognitive Systems Lab

May 2023–Sep 2025

University of Tehran · Research & software

Alongside LLM/RL research, I built an LMS with ChatGPT integration, WebSockets, and educational grouping tools. The platform was tested in two classes and reached over 1,000 users.

Artificial Creatures Lab

Mar 2022–Apr 2024

Sharif University of Technology · Research

Experience-based reinforcement learning, reusable experiment modules, and Unity ML-Agents environments. This collaboration led to the two coauthored papers above.

Quera

Sep 2021–Nov 2022

Teaching · Mentoring · Contest design

I taught and mentored bootcamp participants, maintained frontend courses serving over 3,000 learners, and designed programming contests, including a React recruitment contest with over 1,500 participants.

Beyond the code

a few other interests
Mahyar Riazati

Mahyar / Tehran

I’m just a fully grown, mostly functional computerphile kid.

A compulsive learner with an inconvenient amount of curiosity about the entire computing universe—from high-level abstractions all the way down to PN junctions.

My journey started with electronics, and although software eventually took over, I’ve never been particularly interested in respecting the boundary between hardware and software.

Outside of that orbit: mathematics, theoretical computer science, philosophy, video games, anime, and manga. I have a soft spot for cyberpunk, science fiction, and the questions they raise about technology and what it means to be human.

This is my little corner of the internet. Like most things I build, it’s perpetually a work in progress.

There’s a more informal corner on Telegram .

Elsewhere

links / contact

Keep in touch.

For work, research, or an interesting conversation.

uniq(FIRST_NAME + LAST_NAME − vowels)
lowercase · at gmail dot com