Muhammad Zian

Robot soccer: the orange robot replans its route to the ball as the opponents move.

A* search

Drag an opponent. Click to move the ball.

Computer Science, University of Calgary
Calgary, AB

I build software that makes machines move.

About

Calgary, local time

I've always been fascinated by one question: how do computers actually work?

My name's Muhammad Zian, and I was a curious kid. I was captivated by the idea that a few lines of code could bring a machine to life, execute commands, and perform tasks. That curiosity eventually grew into a passion for robotics and technology.

Growing up, I always envisioned a future in the military. Although life took me down a different path, it led me to Computer Science, where I discovered a way to bring both my interests together.

Today, I aspire to work at the intersection of technology, robotics, and defence: building intelligent systems that solve real-world problems and contribute to something bigger.

For me, Computer Science is more than just writing code. It's about understanding what's possible, turning ideas into reality, and pushing the boundaries of what machines can do.

My goal is simple: to build technology that makes a difference.

Languages
Python, C, Java, JavaScript, TypeScript, SQL, Swift
Robotics
ROS2, Gazebo, grSim, path and motion planning, obstacle avoidance, PID and feedforward control, extended Kalman filter, reinforcement learning
Web and backend
FastAPI, Node.js, Express, React, Next.js, Expo
Data and ML
PostgreSQL, Supabase, pgvector, Redis, RAG, vector search, Gemini, Deepgram
Off duty
Ranked #1 worldwide in Fortnite (C4 S3). Top 10 Radiant in North America in VALORANT (E2 Act 1). Top 250 in Call of Duty: Black Ops 6. Football, soccer and badminton in high school.
1,200+users on a product I founded
3competition podiums in 2026
6developers on the team I lead
6.8%lower estimated pipeline cost in prototype tests

Experience

Roles
Software EngineerAutonomous Intelligent Machines. Motion planning, control and reinforcement learning for robot soccer.Sep 2026 — Now
Technical Product ManagerTech Start UCalgary. Leading six developers building WayLine.Sep 2026 — Now
Quantitative AnalystQuantitative Investment SocietySep 2026 — Now
FounderWCRY. An aim trainer that grew to 1,200+ users.2022 — 2024
Recognition
1st Digital Innovation Case CompetitionFirst in cohort, advancing to the Grand Finals2026 — 27
2nd Cursor HackathonWith CheckpointMar 2026
3rd National DefTech HackathonWith Voice2VectorMar 2026
Education
BSc, Computer ScienceUniversity of CalgaryExpected Apr 2029
BSc, Computer ScienceMount Royal University2025 — 2026
BBA, AccountingMount Royal University2023 — 2025

Work

06 projects

  1. Autonomous Intelligent Machines builds a team of small robots that play soccer on their own. I work on how they move and how they decide.

    Motion planning and control. I use the grSim simulator to measure how the robots respond to movement commands and how accurately they follow a planned path. I have compared PID against PID with feedforward, tested whether an extended Kalman filter improves tracking when the camera data is noisy, and studied when a robot should replan to avoid a collision.

    Reinforcement learning for tactics. I am researching how to design rewards that teach robots when to reposition, pass or shoot, and comparing single-objective with multi-objective learning. One option is a hybrid: a tactical layer weighs scoring chances against defensive risk, and a motion layer executes the chosen action reliably.

    These experiments decide what the team implements and what gets validated on the physical robots. Four of them are rebuilt in the Lab below.

    Role
    Software Engineer
    When
    September 2026 to present
    Built with
    grSim, PID and feedforward control, extended Kalman filter, reinforcement learning
  2. WayLine helps a planner decide where a pipeline should go before anyone commits to a route.

    1. The planner picks a start point and an end point.
    2. WayLine generates candidate routes from terrain, geology, soil, land-use and existing-infrastructure data.
    3. Each route is scored on 6+ constraint categories and on 5 regulatory dimensions: permits, regulators, consultation requirements, approval timelines and regulatory risk.
    4. The routes are compared on an interactive 3D map by construction cost, engineering feasibility and regulatory complexity.

    The idea is regulation-first routing. A shorter route can cost less to build but cross land that triggers extra approvals, and WayLine surfaces that trade-off while the route is still being chosen.

    Role
    Technical Product Manager, leading six developers at Tech Start UCalgary
    Result
    6.8% lower estimated pipeline cost in prototype testing
    When
    September 2026 to present
  3. Drone operators juggle radios, controllers, screens and maps at once. Voice2Vector lets an operator say what they want, such as "take off to 10 metres" or "fly to the northwest watch tower", and keeps their eyes up.

    1. Listen. Speech is captured in the browser and transcribed to text.
    2. Parse. A fast rule-based parser handles standard phrasing. Ambiguous commands fall through to a language model that extracts the intent.
    3. Validate. Every command is checked against the compound layout, no-go zones, altitude limits and what the drone can do. Unsafe commands are rejected with an explanation.
    4. Execute. Valid commands are sent as MAVLink messages to an ArduPilot flight controller, which flies the drone in a Gazebo physics simulation.
    5. Report. The operator sees the result on a 2D map and a 3D view, with every command logged for review.

    A small version of the parse and validate steps runs in the Lab below.

    Result
    3rd place, National DefTech Hackathon, hosted by RCAF and Remote Robotics
    Team
    Four people
    When
    March 2026
    Built with
    React, FastAPI, Python, Gazebo, ArduPilot SITL, MAVLink
  4. Energy companies have to prove they find and repair leaks. Auditors spend hours checking PDF reports by hand against regulatory directives such as D060, D071 and D056. AudItron reads the report, cites the exact rule, and coaches the auditor through each finding.

    1. Upload a PDF audit report and choose which directives apply.
    2. Review. The system flags compliance issues as red, yellow or green. Each one links to the exact regulatory clause and is highlighted on the page of the PDF it came from.
    3. Coach. A voice-first coach asks questions about each finding, so the auditor builds understanding and not dependence.
    4. Document. Every decision is logged with who made it, when and why, which leaves a trail a regulator can follow.

    The hard part was stopping the model from inventing citations, because compliance has no tolerance for that. Every model output has to match a strict schema, and a verification layer rejects any finding that is not grounded in a regulation passage the system actually retrieved.

    Under the hood the directives are split into chunks, embedded as vectors, and searched with a hybrid of keyword and vector search. Long evaluations run as background jobs.

    Built at
    CalgaryHacks 2026
    Team
    Five people
    When
    February 2026
    Built with
    Next.js, FastAPI, PostgreSQL with pgvector, Supabase, Redis, Celery, Gemini, Deepgram, ElevenLabs
  5. Nurses are interrupted constantly, often in the middle of giving medication. Coming back and thinking "did I already give this dose?" is stressful and unsafe. Checkpoint lets a nurse save what they were doing and restore it later.

    1. Capture. One large button records a short phrase, for example "Patel, room 204, metoprolol 25 milligrams, bed alarm".
    2. Structure. The recording is transcribed, and a model extracts the patient, room, medication, dose and notes as structured fields.
    3. Label. Each field is marked confirmed, uncertain or missing, so uncertainty is shown and not hidden.
    4. Resume. A resume screen shows who the patient is, what was being given, what the nurse said, and how long ago.

    Saved checkpoints sit in an unfinished-task queue, and interruptions are logged to a dashboard that shows when and how they cluster. The app also scans patient wristband barcodes. It was my first iOS app.

    Result
    2nd place overall, Cursor Hackathon
    When
    March 2026
    Built with
    Expo, React Native, Node.js, Express, Deepgram, Gemini
  6. I built WCRY in high school because I was playing competitive shooters at a high level and wanted a better way to train. It started as a tool for myself and became a private, invite-only platform.

    AI feedback coached players in real time as they trained. It grew from nothing to more than 1,200 users, and it is what got me into building software.

    Role
    Founder
    Result
    0 to 1,200+ users
    When
    2022 to October 2024

Lab

05 experiments

Bored? These simulations are running live in your browser. Play around and have fun.

01 — Tracking

Kalman filter

An overhead camera tells the team where each robot is, and that data is noisy. Move your cursor over the field and the robot drives towards it. White marks are the raw camera detections, and the orange ring is the filter's estimate of where the robot really is.

Raw error — · filtered —

02 — Control

PID with feedforward

Two identical robots chase the same moving target. The blue one uses PID alone, which only reacts to error, so it trails behind. The yellow one adds feedforward, which tells the controller what the path is about to do.

Average error — PID — · with feedforward —

03 — Measurement

Command response

Before tuning anything, I measure how a robot responds to a movement command. Click anywhere on the field to send the robot there. The chart plots its distance travelled against the command, and the numbers are measured from that run.

Click the field to send a command

04 — Tactics

Pass or shoot

The yellow robot with the ball has to choose. Drag any robot and watch the decision change. The rays show how much of the goal is open, and each pass lane is scored on how safe it is and how good a shot the receiver would have.

This version uses a scoring rule I wrote by hand. My research is about getting robots to learn this trade-off themselves.

Decision —

05 — From Voice2Vector

Command parser

A small rule-based version of the Voice2Vector parser. Type a mission in plain language. Each command is parsed, checked against a 120 m ceiling, the 600 m perimeter and the fuel-depot no-fly zone, and then flown.

    Terminal

    Press / to focus

    Contact

    Working on robots, autonomy, or anything else that has to move in the real world?