Writing

How the desktop app's Local/SSH/Remote sessions quietly replaced a VPS, tmux, and Tailscale setup — and the few cases where a real terminal still earns its keep.

How LLM-based agents observe, reason, and act: model, tools, and orchestration layer, plus ReAct/CoT/ToT with a LangChain example.

Text, image, and multimodal embeddings, vector databases like Pinecone and Chroma, and how ANN search (HNSW, FAISS) makes retrieval fast.

Temperature, top-K, and top-P, plus the prompting techniques that actually move the needle: zero-shot, few-shot, Chain of Thought, ReAct.

From the 2017 Transformer to GPT-4 and Gemini: pretraining, fine-tuning with SFT/RLHF/LoRA, and inference optimization.

Comparing AWS, GCP, and Paperspace GPU instances by VRAM and hourly price for when you hit a CUDA out-of-memory error.

How Glovo cut training-pipeline emissions roughly 100x by reusing prior model state instead of retraining from scratch.

AutoML automates the tedious parts of the ML workflow, but it empowers data scientists rather than replacing them.

From academic tools like Auto-sklearn and TPOT to enterprise platforms like DataRobot and H2O Driverless AI — what open-source AutoML gets you versus paying for it.

Simulating data with the inverse transform and acceptance-rejection methods, from a physics-first take on the Monte Carlo method.

A study of the anomalous magnetic moment of the muon induced by the pseudoscalar-pole contribution to the hadronic light-by-light scattering, using the Dual-Large Number of colors QCD model and Veneziano-type Euler Beta form factors.

Exact diagonalization of cold bosonic atoms under a strong artificial magnetic field in the lowest Landau level, identifying the resulting ground state with the Laughlin state and its fractional quantum Hall characteristics.