Categories

All (57)

TIL (10)

app (3)

daft (1)

dashboard (1)

data (11)

data-engineering (1)

dataframe (1)

db (1)

deep-learning (7)

dependencies (2)

deployment (1)

distributed (1)

docker (1)

fine-tuning (1)

fun (3)

gradle (2)

html (1)

inference (1)

js (1)

jvm (9)

kotlin (7)

llm (1)

loading (1)

machine-learning (20)

mlflow (1)

mlops (1)

multiplatform (2)

nlp (7)

notebook (1)

onnxruntime (1)

packaging (2)

pipeline (4)

presentation (5)

project (1)

pwa (1)

python (11)

raspberrypi (1)

rest (1)

setfit (1)

soft (1)

solara (1)

streamlit (1)

timeseries (5)

type (1)

wasm (2)

web (1)

webapp (1)

workshop (9)

Londogard AB

Read more about us in About.

In here we share our work and exciting side-projects!

Blogs

Transformers.js.py accelerated on-device inference with Python WASM

python

wasm

js

inference

onnxruntime

I’ve previously talked about onnxruntime-web that enables accelerated on-device inference. There’s now a wrapper that proxy Transformers.js to Pyodide. In this post I’ll share how to do just that!

Hampus Londögård

Mini: DeepSeek’s smallpond - a distributed duckdb

python

data

distributed

Open Source software that’d cost millions if not billions of $! I share my quick thoughts on smallpond that runs DuckDB distributed through Ray by DeepSeek (creators of R1 LLM).

Hampus Londögård

Marimo WASM Apps

python

wasm

app

I built a simple but sweet tool to assist you with CSV, JSON & Parquet files. Explore data? Convert? Play around - I got you covered!

Hampus Londögård

TIL: Programatically Fetch Python Class/File Dependencies

python

dependencies

It’s quite common in the Machine Learning realm to launch jobs remotely, and in serverless you essentially do the same. This small gist enables simple tree-shaking fetching (recursively) only relevant local files for your python Class, or file.

Hampus Londögård

MLFlow Models: Self-Contained ML Models with MLFlow

python

mlflow

deployment

machine-learning

Most in the machine learning community knows about MLFlow and uses it to track experiments. One of multiple ‘extras’ is the MLFlow Model which integrates with the Model Registry. The MLFlow Model is quite cool and I’ll share how you can build your self-contained ML models.

Hampus Londögård

Gradio Client - An intro

python

app

rest

machine-learning

Gradio is well-known in the Machine Learning community, and recently they expanded their Apps to automatically include a REST API - Gradio Client, it’s awesome and here’s how it works!

Hampus Londögård

Marimo - A new Notebook/App on the block!

python

app

notebook

Marimo tries to replace multiple tools by being a Reactive Notebook (reproducible), Web App and Python Script at once. It’s a beautiful kit that could end up accelerating in adoption quickly.

Hampus Londögård

Image to Lego (Xmas Project)

python

project

packaging

A childs dream - turning any image into a LEGO Build. In this blog I walk through my xmas project.

Hampus Londögård

Pixi - Real World Usage

python

dependencies

packaging

The last few weeks I’ve put Pixi to the test, utilizing it in real world usage

Hampus Londögård

Data Loading - Comparing Common Tooling

data

loading

Minibenchmark and Developer Experience (DX) between the choices

Hampus Londögård

Data Loading - Daft

TIL

daft

data

The first of multiple blog posts about Data Loading in Deep Learning

Hampus Londögård

ZenML or ClearML? Which MLOps tool strikes best?

mlops

There’s a ton of MLOps tool and I gathered into two open source version that I wanted to try

Hampus Londögård

Streamlit Fragments - Make the Dashboard Dream come true

streamlit

Streamlit is a cool tool that’s incredibly powerful yet simple! Recently they added ‘fragments’ which enables true dashboarding without the continuous full reload!

Hampus Londögård

TIL: Pixi by prefix.dev

TIL

I love micromamba, but pixi seems like the natural next step moving forward!

Hampus Londögård

TIL: Stlite - running streamlit in WASM through b64-encoded URL’s

TIL

I recently learned about Stlite, a port of Streamlit to Pyodide that enables running Streamlit fully inside the browser (WASM)!

Hampus Londögård

TIL: Building Docker Images with Conda and Custom Users (and devcontainers!)

TIL

When building Docker containers there’s a few things to clarify:

Hampus Londögård

TIL: Multiple Git Remotes

TIL

Some instructions on how to use multiple remotes

Hampus Londögård

Solara, League of Legends and Deep Learning to extract E-Sport Highlights

data

machine-learning

deep-learning

fun

solara

I share my journey developing a user-friendly Data App using Solara, where I tackle challenges, gain insights, and implement improvements in threading and Plotly callbacks.

Hampus Londögård

Automatic Highlight Detection of League of Legends Streams

data

machine-learning

deep-learning

fun

In this post I share the way I built a Deep Learning tool to extract LoL Highlights

Hampus Londögård

Solara - ‘A new Reactive Streamlit’

data

dashboard

webapp

Solara is a fresh and exciting web framework that enables React-style state-handling while keeping UI (almost) as easy as Streamlit. Further by design it seems to have a bigger industrial potential than Streamlit.

Hampus Londögård

Kotlin DataFrame vs Polars DataFrame

dataframe

I work using Scala, Kotlin & Python. There’s two DataFrame API’s (polars & Kotlin DataFrame) which I’d like to benchmark.

Hampus Londögård

DuckDB - Quacking Data (WIP)

data

pipeline

DuckDB is particularly well-suited for Online Analytical Processing (OLAP) tasks, which involve analyzing large datasets to gain insights and generate reports. Its columnar storage format and vectorized query engine enable fast and efficient data processing, even on complex analytical queries. Additionally, DuckDB supports a variety of OLAP-specific features, such as window functions, making it a powerful tool for OLAP workloads.

Hampus Londögård

Pandera - Type your Data Pipelines (WIP)

data

pipeline

type

Pandera allows you to decorate functions which will make your DataFrame’s run-time typed. This means that developers & data engineers can expect something to be true, awesome!

Hampus Londögård

Building Actionable Insights from Yelp Reviews using Setfit

deep-learning

llm

fine-tuning

setfit

Setfit: Efficient few-shot learning with Sentence Transformers is a way to train models to perform new tasks with very little data. In this blog I use SetFit to create value for restaurant owners by finding reviews that are helpful and contains improvements applicable to the restaurant using Yelp Dataset.

Hampus Londögård

Automated Data Validation & Exploration

machine-learning

data

A presentation I did at work where we walk through different tools to automate or assist Data Exploration and Validation. A very exciting topic for sure!

Hampus Londögård

Best Practices Are Contextual

soft

pipeline

deep-learning

data

This is something all of us know but sometime forget, or ignore, when we learn about a new technology or methodology in a presentation/blog.

Hampus Londögård

Baby Monitor pt 2

python

I built a baby monitor! Built using Python, Svelte & Raspberry Pi.

Hampus Londögård

Polars - A Refreshingly Great DataFrame Library

data-engineering

pipeline

data

Polars is a DataFrame library written from ground-up to not only have a sensible API but also very efficient operations using multiple cores and clever optimizations such as predicate pushdown & much more!

Hampus Londögård

Probabilistic Forecasting Made Simple

machine-learning

timeseries

While researching probabilistic forecasting in a client project I managed to find a paper which opens the door to any neural network with dropout - which is the majority. That is, we can do probabilistic forecasting with essentially any network!

Hampus Londögård

Timeseries Learnings at AFRY

machine-learning

timeseries

Sharing knowledge based on working a lot with timeseries

Hampus Londögård

Babymonitor #1

python

raspberrypi

I’m building a babymonitor!

Hampus Londögård

Docker (Presentation)

presentation

machine-learning

docker

What is Docker? I went through Docker during our Journal Circle at AFRY X. This is a simple explanation for people not knowledgable about Docker.

Hampus Londögård

GPT2-snapsvisor - Generating Swedish Drinking Songs

machine-learning

nlp

fun

Snapsvisor is traditional Swedish Drinking Songs, sometimes they need some refreshing which I try to do through AI here! ;)

Hampus Londögård

Forecasting Crypto Prices using Deep Learning (Time Series #3)

machine-learning

timeseries

workshop

In this post/Jupyter Notebook we’ll forecast Cryptocurrency prices using Deep Learning (PyTorch, TF/Keras & darts) and we’ll use both both simpler networks and more complex ones like NBEATs.

Hampus Londögård

Predicting Stock Prices using classical machine Learning (Time Series #2)

machine-learning

timeseries

workshop

In this post/Jupyter Notebook we’ll look at stocks, forecasting and predictions using classical machine learning (sklearn) approaches.

Hampus Londögård

Decomposing & Working with Time Series (Time Series #1)

machine-learning

timeseries

workshop

In this post/Jupyter Notebook we’ll look at Time Series and theory surrounding them.

Hampus Londögård

KotlinJS, ONNX and Deep Learning in the browser

kotlin

web

deep-learning

Ever wanted to deploy State-of-the-Art Deep Learning models in the browser? In this blog you’ll learn about how to run inference through onnx webruntime directly inside the browser!

Hampus Londögård

Release nlp (londogard-nlp-toolkit) 1.1.0

nlp

jvm

kotlin

nlp (londogard-nlp-toolkit) has had it’s 1.1.0 release recently with a lot of new functionality and multiple improvements to efficiency, dive in to understand more!

Hampus Londögård

Seam Carving (Presentation & Workshop)

presentation

jvm

kotlin

workshop

Seam Carving is the task to remove empty room in a image. Have you ever wished to do ‘Content Aware Scaling’? Learn it now!

Hampus Londögård

Knowledge Distillation (Presentation)

presentation

machine-learning

What is Knowledge Distillation? How can this allow us to further utilize models and increase efficiency manifold?

Hampus Londögård

TIL: GitHub Pages + Progressive Web App (PWA) = ❤️

pwa

TIL

What are PWAs? How do I create or transform my web app into one? How can I deploy them (freely) via GitHub Pages?

Hampus Londögård

Building a Swedish Named Entity Recognition (NER) model (Flair/Huggingface)

nlp

machine-learning

workshop

Learn how to fine-tune a Flair NER model and quantize a BERT model from Huggingface to achieve SotA performance & a much more efficient model.

Hampus Londögård

When to use what - RegExp, String Replace & Character Replace (JVM/Kotlin)

jvm

A simple quick comparison on which method to use and in what case when replacing characters or strings in strings.

Hampus Londögård

Transformers From Scratch

nlp

machine-learning

deep-learning

workshop

In this post I walk through Self-Attention Transformers from scratch with demos at the end for Text Classification & Generation, where the PyTorch-code is wrapped by fast.ai to simplify end-2-end.

Hampus Londögård

TIL: GitPod - your editor in the cloud

jvm

kotlin

TIL

GitPod is a Cloud IDE where you can run everything from Kotlin to Python/JS. In this TIL how to launch native programs and more is shown in the GitPod IDE.

Hampus Londögård

Object Detection (Workshop)

presentation

machine-learning

How to do object detection using Transformers?

Hampus Londögård

Transfer Learning (Presentation)

presentation

machine-learning

What is Transfer Learning? Can I have the cake and eat it too?

Hampus Londögård

How to build and play Snake via Native Binary, JVM and JS/Browser (Kotlin)

gradle

kotlin

workshop

multiplatform

A three part blog (all included in this one) that goes through 1) How Kotlin Multiplatform works, 2) How to build a game (Snake) and finally 3) how to make it multiplatform.

Hampus Londögård

TIL: SDKMan - The Software Development Kit Manager

jvm

TIL

SDKMan is a tool to make JDK swapping, and installation, simple. It’s really good!

Hampus Londögård

TIL: fastutil - fast & compact type-speciic collections for JVM (no autobox!)

jvm

TIL

Discusses what autoboxing is, why it might hit your performance (and memory). Finally some alternatives are also provided. Learn how to use effective data collections today!

Hampus Londögård

TIL: ‘The Badass Runtime Plugin’, jpackage & jlink - create a ‘native’ installable executable from your JVM-app that isn’t huge

jvm

TIL

‘The Badass Runtime Plugin’ is a plugin that allows you to package a stripped down JRE and modules from your program into a ‘native’ installable program which doesn’t require the user to have Java installed. It takes help of JPackage & JLink to achieve this, my own small program ended up at 35 MB including JRE, which is pretty crazy. This is like Electron, but better! ;)

Hampus Londögård

CoViD-19 FAQ Search Engine 2.0

nlp

deep-learning

machine-learning

workshop

In this post I improve the previous FAQ search engine by some low hanging fruits. The requirements stay the same thus SotA is not achieved but rather it’s simply generic & easy on hardware (Raspberry Pi capable).

Hampus Londögård

SQL - Different Abstraction Levels (& how I came to love SQLDelight)

jvm

kotlin

db

multiplatform

In this post different abstraction levels of SQL is discussed with the final SQLDelight which turns the abstraction into the reverse.

Hampus Londögård

A simple FAQ search engine in Swedish using fastText & Smooth Inverse Frequency

nlp

machine-learning

A search engine for FAQs in Swedish. Completely unsupervised and making use of Word Embeddings & Smooth Inverse Frequency to embed sentences. Basically scratched an itch I’ve had for a while

Hampus Londögård

How I created a email generator in Kotlin (for Afry Tipsrundan)

kotlin

html

This blog goes through how to use Kotlin to generate good looking responsive emails. It’ll handle CSS, kotlin html DSL & kotlin serialization.

Hampus Londögård

Gradle, JVM and GitHub Packages

gradle

jvm

A simple show-case on how to publish libraries to both GitHub Packages & JitPack with a one-button release using gradle & GitHub Actions.

Hampus Londögård

AFRY NLP Competence Meeting: Text Classification IMDB

machine-learning

nlp

workshop

This blog contains my first Competence Meeting where basic NLP concepts where taught and an classifier with good performance was implemented (on IMDB sentiment).

Hampus Londögård