Essays & field notes

Writing

30 entries · 2016—now

ML systems, products and companies—with detours through language, parenting and whatever I’m trying to understand next.

30 entries

2026

Editorial illustration for „I Needed One Obscure Vision Model. Modal Made It Everyone Else’s Problem.“: An obscure vision machine lifted out of a tangled workshop and placed onto a clean remote GPU pedestal, turning a burdensome dependency into an invisible service.

I Needed One Obscure Vision Model. Modal Made It Everyone Else’s Problem.

Agents & AI · Apr 08

Editorial illustration for „Engineering Agents is really UX Engineering“: A human operator and two small, visibly different agent machines sharing one evolving memory object; the visual tension is trust created or broken by state synchronization.

Engineering Agents is really UX Engineering

Agents & AI · Mar 09

Editorial illustration for „The AI-Powered 10-Minute Habit That Taught My Kid to Read (And Made Me a Better Dad)“: A parent and child building a short daily reading ritual from playful image cards, with steady progression shown through physical rhythm rather than charts or interfaces.

The AI-Powered 10-Minute Habit That Taught My Kid to Read (And Made Me a Better Dad)

Life & Learning · Feb 02

2025

Editorial illustration for „Five Practical Lessons for Serving Models with Triton Inference Server“: A compact industrial dispatch station routing several model-shaped modules through one carefully controlled serving lane, emphasizing operational fit over spectacle.

Five Practical Lessons for Serving Models with Triton Inference Server

ML Systems · Dec 15

2024

Editorial illustration for „I’m Not the Founder This App Deserves“: A founder standing outside a beautifully lit product doorway that clearly belongs to another community, choosing fit and identity over an attractive market.

I’m Not the Founder This App Deserves

Building · May 14

2022

LightTag Has Been Acquired by Primer.ai

Building · Feb 15

2021

Editorial illustration for „Complement Objective Training With PyTorch Lightning“: A learner improves by studying the negative space around the correct choice: one illuminated form surrounded by carefully weighted alternatives.

Complement Objective Training With PyTorch Lightning

ML Systems · Jun 23

Editorial illustration for „SnorQL: Scaling Weak Supervision With SQL“: A SQL workbench conducts many weak labeling signals into one scalable stream, preserving the original SnorQL character and concept without logos or text.

SnorQL: Scaling Weak Supervision With SQL

ML Systems · Feb 08

2020

Editorial illustration for „Database Multi-Tenancy for SaaS“: A shared apartment block whose tenants each have a clearly isolated interior behind one maintained facade, expressing schema isolation with domestic precision.

Database Multi-Tenancy for SaaS

Building · Dec 19

Editorial illustration for „Code to Align Annotations with Hugging Face Tokenizers“: Two differently spaced physical ribbons—human spans and machine tokens—being precisely pinned to a shared coordinate rail, editorial rather than instructional.

Code to Align Annotations with Hugging Face Tokenizers

ML Systems · Sep 20

Editorial illustration for „Sequence Labeling With Transformers“: Long annotated fabric is divided into overlapping transformer-sized frames while every marked span stays aligned to its original coordinates.

Sequence Labeling With Transformers

ML Systems · Sep 18

Editorial illustration for „JavaScript String Offsets and Unicode Surrogate Pairs“: One vivid Unicode glyph travels as a paired set of mechanical tiles through a JavaScript counting gate, preserving the original playful mismatch metaphor.

JavaScript String Offsets and Unicode Surrogate Pairs

ML Systems · May 31

Editorial illustration for „ALMa: Active Learning (Data) Manager“: A composed active-learning steward sorting a moving boundary between labeled and unlabeled material, preserving the original manager metaphor with greater editorial clarity.

ALMa: Active Learning (Data) Manager

ML Systems · May 10

Editorial illustration for „Simpledorff: Krippendorff's Alpha on DataFrames“: Several judges place marks onto a shared translucent grid while one calm measuring instrument evaluates agreement without pretending disagreement disappears.

Simpledorff: Krippendorff's Alpha on DataFrames

ML Systems · May 03

Editorial illustration for „Fast NLP Model Pretraining With VAMPIRE“: A compact nocturnal learning engine drawing useful domain signal from a modest local corpus while a giant distant machine remains unnecessary.

Fast NLP Model Pretraining With VAMPIRE

ML Systems · Apr 26

2019

Editorial illustration for „Which Open-Source NER Model Is Best? Comparing CoreNLP, spaCy, and Flair“: Three distinct field instruments inspect the same messy language specimen and disagree in revealing ways, emphasizing practical comparison over a winner podium.

Which Open-Source NER Model Is Best? Comparing CoreNLP, spaCy, and Flair

ML Systems · Nov 20

Editorial illustration for „Context Is King! Why Deep Learning Matters for NLP“: One ambiguous object changes meaning as concentric scenes reveal progressively wider context, with a single strong center and no factual labels.

Context Is King! Why Deep Learning Matters for NLP

Agents & AI · Nov 19

Editorial illustration for „Postgres Range Aggregation for NLP and Everything Else“: Overlapping physical strips nest and merge into maximal containing spans on a clean worktable, conveying range aggregation through material form rather than notation.

Postgres Range Aggregation for NLP and Everything Else

ML Systems · Oct 18

Editorial illustration for „The Joy of IndexedDB for NLP“: A browser-sized cabinet instantly revealing the right document from a dense private archive, celebrating local search with a slightly mischievous physical metaphor.

The Joy of IndexedDB for NLP

Building · Jul 08

Editorial illustration for „How to Label Data“: A careful workshop assembly line turning raw language into dependable labeled material through seven visible stations, with objects and gestures but no text.

How to Label Data

Building · Jun 01

Editorial illustration for „Using dc.js and Crossfilter With React“: Independent chart mechanisms snap into one React frame around a shared filtering axle, expressive editorial machinery rather than a literal dashboard.

Using dc.js and Crossfilter With React

ML Systems · Feb 10

Editorial illustration for „When Should You Use Machine in the Loop?“: A human and machine alternate control of one careful production loop, with visible handoff points and uncertainty rather than an autonomous robot fantasy.

When Should You Use Machine in the Loop?

Agents & AI · Feb 05

2018

Editorial illustration for „Character-Level NLP“: Individual letterforms moving through a long mechanical loom and emerging as semantic fabric, showing the power and cost of character-level processing without embedded text.

Character-Level NLP

ML Systems · Dec 21

Editorial illustration for „Understanding the TensorFlow Estimator API“: A rigid but reusable machine casing accepts interchangeable model parts through a disciplined interface, showing both the structure and friction of the Estimator API.

Understanding the TensorFlow Estimator API

ML Systems · Dec 03

Editorial illustration for „Embrace the Noise: A Case Study of Text Annotation for Medical Imaging“: A radiologist-like observer extracting one reliable signal from layered translucent, noisy chest-film textures; expressive editorial art, not a medical infographic.

Embrace the Noise: A Case Study of Text Annotation for Medical Imaging

ML Systems · Nov 03

Editorial illustration for „Efficiently Labeling Data for NLP“: A labeling workbench that removes repeated, low-value material before a small set of useful specimens reaches the annotator, emphasizing leverage over volume.

Efficiently Labeling Data for NLP

ML Systems · Nov 03

Editorial illustration for „Active Learning: Optimization Is Not Improvement“: A papaya taster repeatedly sampling the most ambiguous fruit while the neglected market remains unevenly understood, separating optimization from real improvement.

Active Learning: Optimization Is Not Improvement

ML Systems · Jun 03

2017

Editorial illustration for „Bug Postmortem: Wrong Image Deployed on Docker Swarm“: Two nearly identical shipping containers cross on a deployment conveyor and one reaches the wrong customer door, a precise visual postmortem with one red anomaly.

Bug Postmortem: Wrong Image Deployed on Docker Swarm

Building · Dec 03

Editorial illustration for „Convolutional Methods for Text“: A sentence transformed through stacked receptive windows, with local fragments progressively joining into a single coherent signal; editorial, tactile, and non-diagrammatic.

Convolutional Methods for Text

ML Systems · May 22

2016

Editorial illustration for „Deep Learning The Stock Market“: A dense market surface compressed into a small memory vessel that carries patterns forward through time, balancing curiosity with the uncertainty of prediction.

Deep Learning The Stock Market

ML Systems · Dec 03