Nino Risteski
I’m a Founding ML Engineer at AIRecruitPro, an open-source contributor, and a self-taught builder who likes learning by making things from scratch.
Most of my work sits somewhere between AI systems, product engineering, and startups. I care about building useful software, understanding the lower-level pieces behind modern ML, and moving quickly from rough idea to something real.
I spend a lot of time around multimodal models, GPU-accelerated training, ML infrastructure, and small tools that remove friction from everyday work.
Lately I’ve been digging into GPU kernels, memory bandwidth, and matrix multiplication performance.
Talks
From Unified Memory to USB GPUs — Apple-First Deep Learning
Talk at iOSKonf26 on Apple-first deep learning, from unified memory architecture to USB GPU setups.
Projects
METAL_SGEMM_M4
Metal compute shader experiments for SGEMM (4x4 tiles) that explore Apple GPU optimization and benchmarking.
TinyPilot
AI agent that helps you learn TinyGrad and keeps you updated on bounties, featuring RAG with ChromaDB and GPT integration.
HubermanAnswers
Ask Dr. Andrew Huberman questions and receive answers directly from his podcast episodes.
smart_screenshots
A tool that renames screenshots immediately based on the context in the screenshot.
Tinygrad
- delete extra cast in decompositions #14517 (Feb 2026) — the extra cast is redundant, bitcast(d.dtype) already returns a float.
- optim empty shard #13513 (Dec 2025) — optimized empty shards so the distributed optimizer skips useless work.
- rope cache optim for jit prune in llm.py (Aug 2025) — cached ROPE calculations so the JIT prune path reuses angles instead of recomputing them.
- adding enable_gqa in SDPA (Jul 2025) — added PyTorch-style grouped query attention plus coverage.
- clean up unused imports in examples and update CI linting (Jun 2025) — cleaned unused imports and wired Ruff so examples stay linted.
- fix test logcumsumexp broken devectorize=0 (Jun 2025) — masked before `exp` and used dtype-aware sentinels to stop NaNs.
- small fix replacing download_file with fetch (Jun 2025) — swapped ad-hoc download glue with the canonical `fetch` helper.
Writing
- RW #9 - How Vector DBs Store 100M Embeddings on One Machine (and Still Search Fast) January 11, 2026 - ML Pills
- You Won't Do Boring Work in The (Near) Future January 3, 2026 - Worklog
- How I Started Contributing to tinygrad — My First 4 Merged PRs August 11, 2025
- Is "Jack of all trades - a master of none" the best mindset for a Data Scientist? January 26, 2023
Useful Stuff
- Job Application StrategyGabriel Peters on standing out with personalized demos
- Reading & ResearchMason Wang on effective reading and research strategies
- School is Not EnoughSimon Sarris on learning through doing and building agency
- Neural Networks: Zero to HeroAndrej Karpathy's practical deep learning course
- Cursor InterviewLex Fridman's conversation with the founding team of Cursor
- Hugging Face CoursesFree ML and NLP learning paths
- Machines of Loving GraceDario Amodei on how AI could transform the world for the better
- The Techno-Optimist ManifestoMarc Andreessen on technology as a lever for progress