RH

uses

my actual stack. one honest line per tool. no sponsored opinions.

languages

Python

default brain. AI, backend, scripts, everything that shouldn't be TypeScript.

TypeScript

JavaScript but with a safety net. Use it for anything that ships to users.

JavaScript

still here. still relevant. use it when TS feels like overkill.

SQL

the language that never left. joins are just philosophy problems with consequences.

C

learned it. respect it. avoid it whenever possible.

ai / ml

PyTorch

primary framework. research-grade control without the masochism of raw NumPy.

TensorFlow

when the dataset has opinions about frameworks.

scikit-learn

for when you don't need a GPU to get a baseline.

Transformers (HF)

because reinventing attention from scratch is a bad use of time.

SHAP

explainability layer. the 'why did the model do that' tool.

llm / nlp

Llama 3.x via Groq

fast, cheap, runs inference at sub-2s latency. the working LLM.

Prompt Engineering

it's not a dark art. it's just being precise. most people aren't.

RAG Pipelines

retrieval-augmented generation — because hallucination is a feature nobody asked for.

backend

FastAPI

Python API server. fast, typed, async-first. the one I reach for first.

Flask

older, lighter, still useful when FastAPI feels like too much ceremony.

Node.js

for TypeScript backends and Next.js route handlers.

Docker

containerise once, argue about environment variables forever.

frontend

Next.js

App Router. the one framework that makes full-stack feel like one thing.

React

components all the way down. sometimes that's a feature.

Tailwind CSS

utility classes, dark mode variables, no style fights with teammates.

Recharts

for data viz that doesn't require a design degree.

Framer Motion

animations that feel like they were meant to be there.

databases

PostgreSQL

the default. relational, reliable, never surprised me badly.

MySQL

when the hosting is opinionated.

Supabase

Postgres with an API. caching SkipTheTerms on this cut costs 40%.

MongoDB

for when the schema is a moving target and you've accepted that.

Redis

in-memory. used for rate limiting. very fast. very ephemeral.

analytics

GA4

11-event funnel on ChannelSpy. 'users visited' isn't analytics.

Microsoft Clarity

heatmaps and session replays. see where users actually click vs where you think they click.

tools & editor

Cursor

AI-first editor. this site was built in it.

Vercel

deploy in 30 seconds. the answer to 'where do I host this' 90% of the time.

Git / GitHub

version control and public proof of work. same thing.

Linux / zsh

terminal-first workflow. tab completion is load-bearing.