Jesper Dr.amsch
@jesper.drams.ch
I share non-hype ML & AI with 8+ years XP 🌦 Machine Learning at ECMWF 💾 Fellow at SSI You miss 99% of benchmarks you don't overfit on! 🏳️🌈 they
💻 pynimate: 359⭐ I needed animated bar chart races and did not want to leave Python for it.
💻 pynimate: 359⭐
pynimate takes a pandas DataFrame with time-indexed data and turns it into animated visualisations -- bar chart races, line plot animations, and more. Pure Python, MIT licensed, pip installable. Nice tool for conference talks or social posts when a static chart does not tell the story.
amplt.de
💻 prompttools: 3 k ⭐ I have been testing prompts by vibes. That is not engineering.
💻 prompttools: 3 k ⭐
prompttools lets you systematically compare prompts across 10+ LLM providers and 7+ vector databases, all running locally. Export results as CSV, JSON, or to MongoDB. There is even a Streamlit playground for non-coders. If you are building anything with LLMs, structured prompt evaluation saves you from yourself.
amplt.de
💻 the-incredible-pytorch: 12.5 k ⭐ I needed a single bookmark for the PyTorch ecosystem.
💻 the-incredible-pytorch: 12.5 k ⭐
The Incredible PyTorch is a curated list covering everything from LLMs and object detection to reinforcement learning, quantization, and medical imaging -- all with PyTorch implementations. Papers, tutorials, libraries, and tools in one place. Solid starting point when you know what you want to build but not which library to use.
amplt.de
I went on The Data Scientist Show with Daliana Liu and we talked way longer than planned. We covered the Kaggle-to-career pipeline, why I switched from Keras to PyTorch, dealing with missing data in geoscience, and what self-taught data science actually looks like when you don't have a CS degree.
I went on The Data Scientist Show with Daliana Liu and we talked way longer than planned.
A good listen if you're figuring out your own path into ML.
amplt.de
💻 tqdm: 31.1 k ⭐ I used to stare at a silent terminal wondering if my script had hung or was still processing. One import fixed that permanently.
💻 tqdm: 31.1 k ⭐
Wrap any iterable with `tqdm()` and you get a live progress bar with percentage, speed, and ETA — 60 nanoseconds overhead per iteration. Works in terminals, Jupyter notebooks, pandas operations, and even as a CLI pipe. Zero configuration needed. If you've ever killed a long-running script because you couldn't tell if it was stuck, tqdm is the cheapest insurance you'll ever add.
amplt.de
Meta keeps open-sourcing what others charge for. Meta is reportedly working on free code-generating AI software, positioning it directly against OpenAI's commercial offerings. The open-source-as-strategy playbook continues.
Meta keeps open-sourcing what others charge for.
The real winners of this AI arms race might be the developers who get free tools out of it.
amplt.de
I built this because I kept missing CFP deadlines by two days. PythonDeadlin.es tracks submission deadlines for Python conferences worldwide -- PyCon, PyData, EuroSciPy, and hundreds more. Filterable by category, exportable to your calendar, shown in your local timezone.
I built this because I kept missing CFP deadlines by two days.
If you've ever rage-googled a deadline you already missed, this exists for you.
amplt.de
💻 machine-learning-notes: 839⭐ Sebastian Raschka's personal ML notes became a public resource.
💻 machine-learning-notes: 839⭐
This repo is a collection of Jupyter notebooks covering hyperparameter tuning, loss functions, learning rate scheduling, regression methods, model evaluation, and more. It started as personal reference material and grew into something genuinely useful for anyone who learns by reading worked examples. Sometimes the best learning resources are the ones someone made for themselves.
amplt.de
💻 bat: 58.7 k ⭐ I've been reading source files with cat for 15 years. Plain white text, no line numbers, no context. bat showed me what I was missing.
💻 bat: 58.7 k ⭐
bat is cat with syntax highlighting, line numbers, git change markers in the margin, and automatic paging for long files. Detects language from the file extension or shebang. Drop-in replacement — aliasing cat to bat is all it takes. If you read code in the terminal and your output is monochrome, `alias cat=bat` is a one-line quality-of-life upgrade.
amplt.de
I left Twitter when it stopped being a place I wanted to share work. Bluesky is where I landed. The ML, Python, and open source conversations actually moved there. Custom feeds mean I see what matters instead of what the algorithm wants me to rage-click.
I left Twitter when it stopped being a place I wanted to share work.
If you're still on the fence: jesper.drams.ch
amplt.de
💻 nicegui: 15.7 k ⭐ I keep wanting Streamlit to be simpler about state, and NiceGUI just is.
💻 nicegui: 15.7 k ⭐
NiceGUI builds web UIs in pure Python using FastAPI + Vue/Quasar under the hood. It gives you buttons, sliders, plots, 3D scenes, and real-time data binding without Streamlit's state magic or having to write JavaScript. Hot reload, Jupyter support, and a pytest-based testing framework included. Great fit for dashboards, internal tools, or any time you want a UI without leaving Python.
amplt.de
Defending a PhD in public is one of the stranger academic rituals. Mine's on YouTube now. The defence covers ML in geoscience: how neural networks can be applied to seismic data, what happens when you build physics into your architecture, and what I learned from four years of research.
Defending a PhD in public is one of the stranger academic rituals. Mine's on YouTube now.
If you've ever wondered what a PhD defence actually looks like, here you go.
amplt.de
💻 aquarel: 860⭐ I spend too much time tweaking matplotlib rcparams. This fixes that.
💻 aquarel: 860⭐
Aquarel is a lightweight theming engine for matplotlib with 13+ built-in themes (gruvbox, solarized, scientific, minimal, and more). Unlike global stylesheets, it uses context managers so you can style individual plots. Themes serialize to JSON for sharing, and transforms handle things like trimmed axes that rcparams cannot. Per-plot theming with a context manager is the detail that sold me.
amplt.de
💻 uv: 83.8 k ⭐ I managed Python environments with pip, virtualenv, and pyenv for over a decade. Then I tried uv and genuinely couldn't go back.
💻 uv: 83.8 k ⭐
uv replaces pip, pip-tools, virtualenv, pyenv, pipx, and poetry — one Rust-based tool, 10-100x faster than pip, with a universal lockfile. It installs Python versions, manages virtual environments, runs scripts with inline dependencies, and even publishes packages. No Rust or Python required to install. If you're still managing your Python environments with multiple tools, the switch is a single install and you'll feel it immediately.
amplt.de
Sixty seconds is apparently enough to explain most ML concepts. I've been testing that theory. Over on @jesperdramsch I break down machine learning ideas into short videos: one concept, no filler. It's a fun constraint that forces clarity.
Sixty seconds is apparently enough to explain most ML concepts. I've been testing that theory.
ML in your scroll without the doom.
amplt.de
💻 Unlearn-Saliency: 145⭐ AI ethics and model weights -- two things that connect more than people realise.
💻 Unlearn-Saliency: 145⭐
SalUn (ICLR 2024 Spotlight) makes models forget specific data by identifying and modifying only the most relevant weights. It achieves near-exact unlearning on CIFAR-10 with just 0.2% gap and works on diffusion models like Stable Diffusion too. As privacy regulations tighten, machine unlearning is becoming essential infrastructure.
amplt.de
💻 git-truck: 674⭐ Bus factor meets data visualization.
💻 git-truck: 674⭐
Git Truck visualizes your git repository's file structure, contributor patterns, and activity hotspots. It runs locally with npx, works offline, collects no data, and is git-provider agnostic. One command and you immediately see who touched what and where the knowledge silos are. If you care about software sustainability (and you should), this makes the invisible risks visible.
amplt.de
Regional ML weather forecasting got real. Met Norway extended AIFS with a stretched grid that concentrates 2.5 km resolution over the Nordics while keeping a global context. Trained on ERA5 plus just 3.3 years of regional data, it outperforms MEPS for 2m temperature.
Regional ML weather forecasting got real.
Data-driven models can go local without losing the global picture.
amplt.de
💻 fastapi: 96.9 k ⭐ I built REST APIs in Flask for years — route decorators, manual request parsing, Swagger docs as an afterthought. FastAPI made all of that automatic.
💻 fastapi: 96.9 k ⭐
FastAPI builds APIs from Python type hints — request validation, response serialization, and interactive docs (Swagger + ReDoc) all generated from your function signatures. Async-native, performance comparable to Go and Node.js, powered by Starlette and Pydantic under the hood. If you're starting a new API and reaching for Flask out of habit, try FastAPI for one endpoint. You'll notice the difference immediately.
amplt.de
I write about ML, Python, and occasionally zombies. It's a range. The blog covers practical stuff like pytest patterns and clustering in Python alongside weirder pieces on the uncanny valley and AI vacation planning. No posting schedule, just whenever something bugs me enough to write it down.
I write about ML, Python, and occasionally zombies. It's a range.
For when you want data science content that doesn't take itself too seriously.
amplt.de
this is a good article. we are changing what parts of coding are hard. but make no mistake, there are still hard parts. And we need to not gloss over those.
Jesper Dr.amsch @jesper.drams.ch · 6mo agoI look around on LinkedIn and people want AI to do all the hard work! 😅🤖 Here’s a great read on Stack Overflow on why Generative AI can't replace human engineering teams by Charity Majors.
💻 infinite-fractal-stream: 30⭐ What if your training dataset was literally infinite?
💻 infinite-fractal-stream: 30⭐
Infinite-fractal-stream generates procedural fractal images on-the-fly using parameterized Mandelbrot variants, implemented in Triton for GPU speed. No fixed dataset size -- just infinite unique samples for training classifiers, diffusion models, or VAEs. Configurable resolution, depth, and class counts. As ML shifts toward scaling on more data, benchmarking against finite datasets starts to feel like the wrong test.
amplt.de
💻 picklescan: 397⭐ Every ML model you download as a pickle can run arbitrary code. That should concern you.
💻 picklescan: 397⭐
PickleScan detects malicious globals in Python pickle files -- the kind that execute code during deserialization. It scans local files, URLs, zip archives, PyTorch models, numpy .npy files, and Hugging Face repos. ClamAV-style exit codes make it easy to integrate into CI pipelines. If you load untrusted model weights, this belongs in your workflow. Hugging Face already uses it.
amplt.de
I made a course about AI art because I wanted to demystify the part most tutorials skip. It covers Stable Diffusion and prompt engineering on Skillshare -- not just "type this and get that" but understanding why certain prompts work and how the model actually interprets your words.
I made a course about AI art because I wanted to demystify the part most tutorials skip.
For when you want to make AI art with intention, not just luck.
amplt.de
💻 hypothesis: 8.5 k ⭐ I was writing unit tests with hand-picked inputs for years. Hypothesis found a bug in the first function I pointed it at.
💻 hypothesis: 8.5 k ⭐
Hypothesis generates hundreds of random inputs within ranges you define, checks that your assertions hold for all of them, then shrinks any failure to the minimal reproducing case. You write one test instead of twenty, and it finds edge cases you'd never think to check — empty strings, boundary values, Unicode weirdness. If your tests only cover the cases you thought of, that's exactly the problem hypothesis solves.
amplt.de
💻 weatherbenchX: 196⭐ This is literally what I work on, so I have opinions.
💻 weatherbenchX: 196⭐
WeatherBenchX is Google's next-gen framework for evaluating weather forecasts, succeeding WeatherBench 2. Built on xarray, modular by design, and scalable via Apache Beam. It handles sparse station data and satellite observations, not just gridded fields -- which is where real-world evaluation gets hard. If you evaluate ML weather models, this is the benchmarking infrastructure to watch.
amplt.de
💻 The-Little-Book-of-ML-Metrics: 995⭐ I still look up the difference between macro and weighted F1. No shame.
💻 The-Little-Book-of-ML-Metrics: 995⭐
NannyML's Little Book of ML Metrics covers evaluation across regression, classification, clustering, ranking, CV, NLP, GenAI, probabilistic models, bias/fairness, and data observability. Free digital version, open source. Goes from accuracy to the obscure ones you forget exist until you need them. A reference that covers metrics you use daily and the ones you will eventually need.
amplt.de
This is literally my day job and it still blows my mind sometimes. ECMWF now shows ML weather forecasts from FourCastNet, Pangu-Weather, and GraphCast alongside the operational IFS on our public charts. Same verification, same standards. Plus we open-sourced ai-models so anyone can run them.
This is literally my day job and it still blows my mind sometimes.
You can compare physics-based and ML forecasts yourself, right now.
amplt.de
💻 textual: 34.9 k ⭐ I wanted a quick UI for a Python tool but didn't want to learn Qt or ship an Electron app. Textual let me build it in the terminal.
💻 textual: 34.9 k ⭐
Textual is a TUI framework with CSS-like styling, a widget library including data tables, tree views, and input forms, plus a command palette out of the box. Apps run in the terminal or can be served in a browser with no code changes. From the team behind Rich. If you've ever wanted to build an interactive Python tool with more than print statements but less than a full GUI, this is the sweet spot.
amplt.de
💻 pydantic: 27.2 k ⭐ I used to write manual validation for every dict that came from an API or config file. Pydantic made me wonder why I ever did that.
💻 pydantic: 27.2 k ⭐
Define a Python class with type hints, pass in data, and pydantic validates, coerces, and structures it automatically. Strings become datetimes, nested JSON becomes typed objects. V2 is a ground-up Rust-backed rewrite — faster, stricter, and still backward-compatible with V1. If you're still validating data with if-statements and try/except blocks, pydantic will save you thousands of lines.
amplt.de