Hi, my name is

Kiyarash Fazeli.

Senior Software Engineer

Dev Team Lead at Maktabkhooneh, based in Muscat, Oman. I work on Django backends, search and recommendations, and contribute to open-source machine learning libraries.

About Me

Software engineer at maktabkhooneh.org for over ten years. I’ve worked on the Django backend, search, recommendations and the move to Kubernetes. I now lead the dev team.

I organize the Tehran Python User Group (2,000+ active members) and give regular talks there on using ideas from CS research in everyday engineering work.

Outside work I enjoy sitcoms (especially Seinfeld), World War II history, discrete math and cryptography, and applications of logic in computer science.

I go by Kiyarash; Mohammad is the name on my passport.

Skills I work with most:

Machine Learning
Natural Language Processing
Python
Django
Kubernetes
Linux

Experience

Maktabkhooneh

Dev Team Lead - Maktabkhooneh

2023 – present

I lead the backend and frontend engineers and run sprint planning and bug triage; my focus is API performance, Core Web Vitals, error tracking with Sentry, and the move to Nuxt 4.

SRE Team Lead

2020 – 2023

Rewrote search ranking, built a personalized-PageRank recommendation engine, rolled out CI/CD across the company, moved all services to Kubernetes and redesigned the interview process.

Senior Software Engineer, DevOps

2018 – 2020

Rebuilt disaster recovery for video delivery, spread infrastructure across two availability zones, and built an internal tool that automates the video edit-and-publish pipeline.

Software Engineer (part-time)

2015 – 2018

Built the Android app and migrated the backend to Django while studying for my B.Sc. and M.Sc.

Open Source

Python Array API

scikit-learn

Made LinearDiscriminantAnalysis work with array API inputs whose namespace and device differ, such as mixed NumPy and PyTorch/GPU arrays (merged).

Pull request

Python Online ML

river

Added preprocessing.GapEncoder, a streaming version of skrub’s GapEncoder that maps messy categorical strings to latent topics one sample at a time; a worked example is in review.

Pull requests

Python Cython Recommender systems

implicit

Logistic matrix factorization (LMF) was quietly broken: its negative sampling barely showed the model any negatives. I found and fixed four bugs, including a clobbered loop variable that drew one negative instead of neg_prop × seen items and a loop bound tied to the embedding width instead of the catalogue size. Five regression tests now cover them (in review).

Pull request

Python PostgreSQL

BeanQueue (bq)

Made workers exit promptly on Ctrl-C and other kill signals: the metrics and worker-update threads now get a thread-safe stop signal instead of being joined with a timeout and killed.

Pull requests

Python NLP

spaCy

Fixes for quadratic regex backtracking in URL matching, NER format autodetection, and the Greek and Portuguese affix rules (in review).

Pull requests

Projects

Python spaCy NLP

spacy-persian

Persian (Farsi) NLP pipeline for spaCy: named entity recognition, part-of-speech tagging, dependency parsing and lemmatization. The transformer pipeline beats Hazm’s published BERT dependency parser on the same treebank (LAS 90.79 vs 89.34, UAS 93.87 vs 92.46), with 97.6% POS and 97.3% lemma accuracy.

GitHub

Python PyTorch spaCy

floret-torch

GPU training for floret (fastText + Bloom subword table) vectors in PyTorch, exporting spaCy-compatible .floret and .vec files. About 20× faster than CPU floret, and the vectors are just as good (WordSim-353 0.4740 vs 0.4738).

GitHub

Python spaCy Transformers

Persian sentiment analysis (SnappFood)

spaCy text classifiers for SnappFood food-delivery reviews. Fine-tuned ParsBERT reaches 87.3 macro F, just behind the best published results (ParsBERT and mBERT, 87.9–88.1). A floret-vector CNN (86.2) beats the published ALBERT-fa fine-tune and every frozen-embedding result.

Model

Python pandas