sepehr.Resume
Machine learning engineer · Applied researchSan Jose, California

A few thingsI’ve taughtmachines.

I’m Sepehr. I build and study learning systems,
from game-playing agents to medical imaging.

Explore the work
A sculptural metal track with an orange glass bead poised at a fork, on an ivory ceramic base.
+Fig. 00 · A bead at a fork
Sepehr SeghatoleslamiPortfolio / 2026

01 / About

The human bit.

I started with olympiad problems.
Now the experiments run in Python.

I studied computer science at Western University. The projects here run from learning agents and medical imaging to retrieval and game theory.

B.Sc. Honours Specialization in Computer Science
Western University · 2026

Apr 2026 — Present

Research Assistant

Still investigating

Cognitive Development and Neuroimaging Laboratory
Western University

Developing computational tools for neuroimaging research. One of my early projects was an automated MRI segmentation pipeline.

2020 — 2024

Mathematics & CS Tutor

Self-employed · Iran & Canada

Olympiad combinatorics and geometry. University calculus, linear algebra, and computer science.

Grades and medals

92%

Major average

89%

Overall average

Dean’s Honor List
Western Admission Scholarship

2021

Silver medalIranian National Mathematical Olympiad

2020

Bronze medalIranian National Mathematical Olympiad

2021

Third prizeEuropean Mathematical Cup

Usually within reach

Vision

PyTorch

CNNs & segmentation

Robustness evaluation

Learning agents

PPO · DQN · Q-learning

Stable-Baselines3

Gymnasium

Retrieval

FAISS

DuckDB · Polars

Ranking & LLM evaluation

Systems

Python · C/C++

Parallel simulation

Git · Linux

02 / Selected work

Things I’ve built.

Six projects, roughly
in the order I’d bring them up.

01 / Information retrieval
Layered translucent replay panels in a brushed-aluminium frame.

VODLens CS2

Learn from
the replay.

Comparable pro plays. Video evidence attached.

1.93M situationsPython · Retrieval · Evaluation
02 / Undergraduate thesis
0.852 / 1Normalized profit gain · published reference 0.849
CompetitionMonopoly

Algorithmic pricing

When rivals
get along.

Two agents learn to price. An audit tries to catch them getting along.

1,000 learning sessionsQ-learning · Game theory
04 / Control in simulation

88.7%

Mean time in range / virtual adults

PPO88.7%
PID59.0%
Basal-bolus¹95.2%

¹ Receives meal information unavailable to PPO.

AutoDoseRL

Learning to
stay in range.

A dosing policy with explicit safeguards.

10 virtual patientsPPO · Continuous control

Further work

05—06

05 / Visual robustness

A change of scenery.

Ceramic and metal track before transformation.
Original
blur transformation of the same composition, intensity 45 of 100.
Blurred

Turn it up45%

Try a transformation. The study measured three training recipes under 50 of them.

See what held up

06 / Poker & detectability

The price of
a poker face.

≥ +1.95

Mixed honest agents

↔
≥ +0.51

Selfish CFR agents

One detector, calibrated against two ideas of honest play. Best-found values, in chips per hand.

See what stays hidden