me, in glyphs
team lead at adigy.ai · almaty
Hi, I’m Temirlan. I make machine learning boring enough to trust.
I’m an AI engineer, and I like building things. I’ve worked across web development, machine learning and team leading, in consulting, robotics, ads management, medical AI and several e-commerce platforms.
fig. 0 — draw a digit, a neural net reads it97.2% on 10,000 test digits
draw a digit here
what the network sees · 28×28
64 hidden neurons
it’s a–
A 784 → 64 → 10 network I trained on MNIST in plain NumPy (train script). Weights are int8, 68 KB, and run in your browser.your drawing never leaves the page
+20%user revenue from a new bid-optimization systemadigy.ai
−35%infrastructure costadigy.ai
−90%labeling cost with active learningsewts
2peer-reviewed papersresearch
Selected work
08 · live in your browserThe lab
real models, running in your browserfig. 1 — photo → design tokens, by k-meanspick a photo or use your own
clusters → tokens · iteration 0
new
Your product, re-themed
Every colour here comes from the photo. Roles are picked by lightness and saturation.
Accessible contrast
k-means++ on 4,000 sampled pixels in Lab colour space, k = 5. Watch the centroids settle, then WCAG checks on the roles.
fig. 2 — ask my resumebm25 retrieval, the “r” in rag
⌘K
Okapi BM25 (k1 = 1.2, b = 0.75) over lines of my resume, with light stemming and a small synonym map. Highlighted words matched.
fig. 3 — three optimizers
One valley, real gradients. SGD, momentum and Adam race down the Rosenbrock function from the same start.
sgdmomentumadam
hint
Click anywhere on the landscape to drop a new start point. Drag these windows around.
fig. 4 — active learning, liveorange = still wrong
A real classifier learning in your browser. It asks for the digits it is least sure about. Hover to label some yourself.
fig. 5 — minhash, the trick inside traceredit either text
128 hash minimums over word 3-grams. Matching slots estimate the overlap without comparing the texts.
Experience & stack
2020 — nowSix years, six teams. From full-stack platforms to ML pipelines for robots and ad bidding.
PythonTypeScriptPyTorchLightningHugging FaceLangChainLlamaIndexscikit-learnFastAPIDjangoNext.jsReactPostgreSQLRedisFAISSDockerTerraformAWSGCPDatadog
Experience2020 — now
Software Engineering Team Lead10.2025 — nowAdigy.ai · Amazon ads management · remote, USA
- State-management system for bid optimization with market-weighted averages: user revenue +20%.
- Reworked infrastructure and queries with AWS and Datadog: monthly cost −35%.
- Introduced TDD, test coverage 4×. Release cycle from 10 days to 3.
ML & Software Engineer08.2024 — 10.2025Sewts GmbH · robotics and automation · Munich
- Containerized PyTorch pipelines with continuous retraining on live data: +10% accuracy.
- Active learning in the annotation loop cut labeling cost by 90%.
- Diffusion model for synthetic textile images doubled the dataset.
Software Engineer / Team Lead09.2023 — 06.2024Fleek.kz · logistics platform · remote
- Logistics platform on FastAPI and Next.js with route optimization and shipment tracking.
- MVP in 2 months, twice as fast as planned. Scrum Master, onboarded juniors.
Software Engineer12.2022 — 08.2023Macdent.kz · Astana
- Query optimization and predictive caching: responses 10× faster.
- Moved to AWS for 5× faster loads; backups, sharding and CI/CD.
Machine Learning Researcher08.2021 — 11.2022Nazarbayev University · Astana
- Two peer-reviewed papers in applied ML and computer vision.
- Retraining pipeline for mobile pest detection: model +50%. YOLOv5 on edge devices.
Full-Stack Software Engineer07.2020 — 08.2021BRI · courses platform · Astana
- Mobile version of the product: daily active users 20k → 40k.
- Partner integrations: customer retention +50%.
Publications2
Enhancing Machine Learning Training Performance in Smart Agriculture Datasets Using a Mobile App2023 ↗IEEE MetroAgriFor · Zarymkanov, Kargar, Pinotti, O’Flynn, Zorbas A Machine Learning Model Selection Considering Tradeoffs between Accuracy and Interpretability2022 ↗ASTES Journal · Nazir, Zarymkanov, ParkEducation2
MSc Data Analytics2023 — 2025University of Hildesheim, Germany · ML, computer vision, optimization
BSc Computer Science2019 — 2023Nazarbayev University, Kazakhstan