Stack

I spend too much time optimizing this. Here's what I actually use.

Daily Stack

Everything here earns its place.

Cursor — AI-native IDE. Tab-completions that actually work. I don't think about it anymore.
VS Code — Side-by-side with Cursor for extensions it hasn't caught up on yet.
Warp — Terminal with natural language command suggestions. Saved me on flags I use twice a year.
Raycast — Replaced Spotlight entirely. Launcher, clipboard history, and custom scripts in one shortcut.
Notion — Project wikis, research notes, anything that needs structure and sharing.
Obsidian — Local-first thinking. Fast, offline, plain text, and not going anywhere.
Claude API — Default for inference and most LLM work. Hard to argue with the context window.

Things I'm experimenting with but haven't fully committed to.

Linear — Moving bigger projects here from Notion. Early impressions are very good.
Perplexity — Research queries where I want citations, not vibes.
Granola — AI meeting notes that actually capture what was said, not just a transcript.
Wispr Flow — Speech-to-text that works system-wide. Cuts writing time on long docs significantly.

Core toolchain for most of what I build.

PyTorch — Default framework. Rewards understanding what's actually happening under the hood.
Hugging Face Transformers — Model hub and pipelines. Rarely write model code from scratch anymore.
Weights & Biases — Experiment tracking. Ran 50 experiments without it once — never again.
LangChain — Fast LLM pipeline prototyping. Replace with raw API calls before anything goes to production.
FastAPI — Serve a model in 30 lines. Go-to for endpoints.
Lambda Labs / Vast.ai — Rent GPU compute instead of buying. Cheaper and more flexible than owning hardware that depreciates.

Research-specific tools. Underrated stack for the intersection of ML and life sciences.

RDKit — Molecular fingerprinting, SMILES parsing, descriptor generation. Best cheminformatics library by a wide margin.
BioPython — Sequence analysis and database interfacing (NCBI, UniProt). Saves hours on data wrangling.
DeepChem — ML models that understand molecular graphs out of the box. Great for baselines.
AlphaFold / ESMFold — Protein structure prediction. Changed what's possible in structural biology research.
AutoDock Vina — Docking simulations. Core to my work on ligand-receptor topology and drug discovery.
Mol* — Browser-based molecular visualization. Reach for this before firing up PyMOL.
PDB & PubChem — Open databases I have open all day, every day.

What I run code on.

MacBook Pro M2 Pro (14-inch) — Handles everything short of serious training runs. Battery life is genuinely good.
Raspberry Pi 4 — Edge inference prototyping and anything physical. Keeps a permanent spot on my desk.
Arduino Uno — Lower-level hardware interfacing when the Pi is overkill. Great for sensors and actuators.
NVIDIA Jetson Nano — Runs lightweight models at the edge without a cloud round-trip. Useful for latency-sensitive demos.

How I consume information without drowning in it.

ArXiv (cs.LG, cs.AI, q-bio.QM) — Daily digest. Filter aggressively or drown.
Papers With Code — Reproduction baselines before I build on top of anything. Saves a lot of wasted effort.
Readwise Reader — Long-form reading that surfaces highlights back at the right time.
Twitter / X — Real-time ML discourse. Signal-to-noise ratio is genuinely terrible. The signal is still worth it.