Now / 2026-08-28
About explains the stable thread. This page records the current state of the journey, so it can change without changing the whole identity.
Last updated · 2026-08-28
01 / Learning
Retrieval, agents, evaluation, and the decisions that make an AI feature useful beyond a demo.
02 / Building
Full-stack web work, data experiments, and small systems that turn understanding into something real.
03 / Exploring
Communication, community, public learning, and collaborations that help technical work travel further.
04 / Questions
How can I show the reasoning behind a result? What should remain human?
05 / Learning
The order matters: strengthen the mathematical foundation, deepen computer vision, then connect models to the systems that make them useful in the real world.
Understand what happens inside a model mathematically, not only how to call YOLO or PyTorch.
Neural networks · backpropagation · gradient descent · loss · optimization · CNNs · attention · transformers
Goal: read a new equation or architecture and understand why it works.
Deepen the specialization already connected to my real projects, from detection and tracking to evaluation and deployment.
IoU · NMS · precision/recall · mAP · tracking · augmentation · domain shift · quantization · CUDA · TensorRT
Goal: connect the mathematics of vision to measurable system behavior.
Study the architecture behind modern language models and the systems that make inference efficient.
Self-attention · positional encoding · KV cache · tokenization · embeddings · fine-tuning · LoRA/QLoRA · vLLM · PagedAttention
Goal: understand both the model and the memory/computation decisions around it.
Build the bridge between software engineering, AI, and systems: from a model and PyTorch to serving, monitoring, and profiling.
GPU architecture · CUDA · memory · kernels · batching · async inference · concurrency · serving · latency vs throughput
Goal: make AI systems faster, clearer, and more reliable—not just more impressive.
Train myself to ask better questions: what problem remains, which assumptions matter, and what evidence would change the conclusion?
Problem · prior work · limitation · idea · formulation · architecture · experiments · ablations · results · limitations
Goal: move from using models to reasoning about what could be improved.
My next focus is to turn one of these explorations into a properly documented project: a working artifact, a clear explanation, and a record of what I learned while building it.
Explore the work ↗