Python on sysid blog

How Claude Code Saved My Weekend

Sat, 26 Jul 2025 00:00:00 +0000

If you want something done right, automate it.

Modern Open-Source Maintenance

As the maintainer of sse-starlette, my GitHub notifications rarely bring me joy on a lazy Saturday afternoon.

This weekend, I decided to push Claude Code a little bit, hoping to save precious summer time.

Issue #140 didn’t just look interesting; it appeared non-trivial, touching on asyncio intricacies, multi-threading, and event-loop handling—elements that typically guarantee headaches.

Rather than my usual deep-dive session of debugging and refactoring, I handed the reins to Claude Code, eager to see if it could manage the problem.

Anki, Markdown, and the Power of inka2

Sun, 10 Mar 2024 00:00:00 +0000

Streamlining Knowledge Management: Anki, Markdown, and the Power of inka2

Introduction

Acquiring and keeping knowledge is an uphill battle. Luckily the digital age has awarded us two powerful tools: Anki for spaced repetition learning, and Markdown for its straightforward syntax in note-taking. Bridging these worlds is inka2 a tool designed to transform Markdown notes into Anki flashcards with seamless efficiency.

Why Markdown for Knowledge Management?

Markdown stands as the go-to format for many developers and writers, praised for its simplicity and versatility. It allows for quick note-taking, organizing thoughts, and even documenting complex code snippets without breaking a sweat. Coupled with a text editor like Vim, managing knowledge becomes a streamlined process, minimizing distractions and maximizing productivity.

Linear Optimization in time-critical Control Loop

Sat, 13 May 2023 00:00:00 +0000

Problem

Given a start and endpoint, find the eager and lazy path from start to end.

Model

The visualization translates to a simple rhombus (German: Raute) where the enclosed area is to be maximized.

The task is to find the corner points A, B, C, D, of the Raute.

There is a twist, however. The points must be contained within a “valid” area, whereas start and endpoint may lie outside.

Select Points

Sun, 30 Apr 2023 00:00:00 +0000

Given multiple sets of data points. Select one point per set so that the distances between the points is minimal.¹

Model

Non-convex MIQP Model

$$ \min \sum_{i,j| ok_{i,j}} dist_{i,j} \cdot x_i \cdot x_j \\ \sum_{i|group_{i,g}} x_i = 1 \forall g \\ x_i \in {0,1} $$

Linear MIP Model

To linearize the model a standard procedure can be applied:

Let’s introduce a new binary variable, $z_{i,j}$, that represents the product of the binary variables $\color{darkred}x_i$ and $\color{darkred}x_j$. This new variable will replace the nonlinear term in the objective function, making the model linear. We will also add constraints to link the new variable $z_{i,j}$ with the original variables $\color{darkred}x_i$ and $\color{darkred}x_j$.

Software Architecture 2022

Fri, 31 Dec 2021 00:00:00 +0000

A refresher in software architecture philosophies with Perseverance, Patience and Python.

This is a high-speed overflight over current approaches with an occasional deep dive.

CRUD, MVC

simple concepts and therefore effective communication in dev teams
works as expected. Limits known, e.g. modelling complex business domains

DDD and Clean Architecture

structure of software logic reflects business processes
ubiquitous language as common communication ground for developer and business
concepts like Aggregate and UnitOfWork provide clear consistency and concurrency model. Facilitates communication.
Clean Architecture and DI is conducive to structured and decoupled code (SRP)

Every system has these two kind of arrows: the dotted-green one showing the flow of control (or the flow of data in your system) and the red one showing the source code dependency.

Maximal Overlap of Rectangles

Sun, 07 Feb 2021 00:00:00 +0000

Find the maximal overlap of rectangles so, that every selected rectangle covers the full width of the overlap area.

A selected rectangle can be larger than the overlap area, but not smaller.

Problem Parameters:

a set of rectangles aligned on the x-axis
rectanges can have different heigt
rectangles show arbitrary overlap

Model

Parameters

$$ i \in [1..N],\ \text{ index: N rectangles }\\ s_i \in R^+,\ \text{ start of rectangle $i$ }\\ e_i \in R^+,\ \text{ end of rectangle $i$ }\\ h_i \in R^+,\ \text{ height of rectangle $i$ }\\ M \in R^+,\ \text{ Big-M constant, upper limit for x }\\ $$

Scheduling

Wed, 06 Jan 2021 00:00:00 +0000

I already analysed several scheduling problems:

The class of scheduling problems is interesting and every example provides new insight. Here I am going to look at two problems which can be solved with a very similar and quite common approach: The continuous time model approach.

Problem 1: Machine Scheduling

Schedule jobs onto available machines so, that the total processing time (a.k.a. makespan) is minimized.

Inventory Management

Fri, 01 Jan 2021 00:00:00 +0000

Find the best order policy to replenish stock levels, i.e. minimize the total cost incurred.

The (Q, R) inventory model is as follows:

If the inventory falls below threshold R, place an order of quantity Q.
Lead times: when an order is placed it takes some time to be delivered
When inventory becomes zero, additional demand will cause a backlog. Backlogged demand will be fulfilled when replenishments arrive but at a cost.

Cost Contribution

Fixed ordering cost (Fixed cost)
Holding cost related to inventory (Holding cost)
Penalties related to backlogs (Shortage cost)

Choose quantity Q and ordering threshold R to strike a good balance among these three costs!

Job Shop Scheduling Problem

Sun, 02 Aug 2020 00:00:00 +0000

The Job Shop Scheduling Problem (JSSP) is an NP-hard problem. ¹

The JSSP is defined by a set of jobs that must be executed by a set of machines in a specific order for each job. Each job has a defined execution time for each machine and a defined processing order of machines. Also, each job must use each machine only once. The machines can only execute a job at a time and once started, the machine cannot be interrupted until the completion of the assigned job.

Resource Constrained Project Success Plan

Sat, 18 Jul 2020 00:00:00 +0000

Scheduling tasks in a large project translates to facing a graph optimization challenge:

The Resource-Constrained Project Scheduling Problem is a combinatorial optimization problem that consists of finding a feasible scheduling for a set of $n$ jobs subject to resource and precedence constraints.

Problem Domain: Network Optimization.

Taks are represented by nodes. Dependencies of tasks on predecessors form a directed graph. Every task has got a duration which is represented by the weight of the directed edges of the graph. Tasks consume limited resources. In order to provide an entry and exit point to the graph we introduce two dummy tasks, $x_0, x_{n+1}$ with duration 0 and no resource consumption (a.k.a. sentinels).

Installing CBC on MacOS

Wed, 24 Jun 2020 00:00:00 +0000

Update May 2022: Using the coinbrew approach as documented here works now out of the box.

Today I needed to install CBC on MacOS and it was not a walk in the park, unfortunately.

Yes, there is a brew package which makes installation trivial. However, this only installs a single-threaded version of CBC. Not good!

One of the greatest selling points of CBC is multithreading. This makes CBC a contender for serious optimization tasks in MIP programming.

Dispersion

Sun, 21 Jun 2020 00:00:00 +0000

Find 10 points out of 50 with maximum distance from each other.¹

This problem can modelled as a max-min model where we maximize the smallest distance of 10 points.

Model

$$ \max \Delta\\ s.t.\\ \Delta \le d_{i,j} + M(1- x_i x_j), \forall i \lt j \\ \sum_i x_i = k\\ x_i \in {0,1}\\ $$

M is an upper bound on the distances: $$ M = \max_{i\le j} d_{i,j}\\ $$

The quadratic part $x_i x_j$ of the constraint can be linearized:

The Book of Mormon II

Sat, 13 Jun 2020 00:00:00 +0000

The full version of the problem you can find in part I of the article.

Recap Challenge

We want to read the book in a given number of days: 128.
We want to read an integer number of chapters each day (there are more chapters than days), and at least 1 chapter each day.
The chapters are very non uniform in length (some very short, a few very long, many in between) so we would like to come up with a reading schedule that minimizes the variance of the length of the days readings (read multiple short chapters on the same day, long chapters are the only one read that day).
We want to read through the book in order (no skipping ahead to combine short chapters that are not naturally next to each other)¹.

Objective:

We minimize the variance of the number of verses read per day.

The Book of Mormon

Sun, 07 Jun 2020 00:00:00 +0000

The Book of Mormon is a sacred text of the Latter Day Saint movement, which, according to adherents, contains writings of ancient prophets who lived on the American continent from approximately 2200 BC to AD 421.¹

Not the easiest matter, so we need a smart reading plan.

Challenge

We want to read the book in a given number of days: 128.
We want to read an integer number of chapters each day (there are more chapters than days), and at least 1 chapter each day.
The chapters are very non uniform in length (some very short, a few very long, many in between) so we would like to come up with a reading schedule that minimizes the variance of the length of the days readings (read multiple short chapters on the same day, long chapters are the only one read that day).
We want to read through the book in order (no skipping ahead to combine short chapters that are not naturally next to each other)².

Yet Another Math Programming Consultant analysed various solution approaches. I want to replicate here two of them with Pyomo as modelling language.

String Puppets

Mon, 01 Jun 2020 00:00:00 +0000

import mona1 from “./mona1.mp4”

I am a fan of the Augsburger Puppenkiste. Puppets are brought to live by pulling strings. If you ever get to Augsburg, do not miss to see a show!

Today a came across an excellent article ¹, which creepily reminded me of these happy String Puppets.

I decided to hit two birds with one stone: Learn something about the current state of Deep Fakes and try out Google’s Colab environment.

Numberlink Puzzle

Sat, 30 May 2020 00:00:00 +0000

Problem

Connect each two cells with the same numbers by a line.
The sections of a line run horizontally or vertically.
Each cell must be visited exactly once by a line.

Solution

The trick is to identify endpoints of numberlinks. They do have only one neighbor with the same value. Cells wich are part of a path must have two neighbors with the same value.

The challenge is now to encode this insight into a Mixed Integer Programming model and fire up a solver.

Server Sent Events

Wed, 27 May 2020 00:00:00 +0000

If you are looking for realtime features in WEB applciations you often end with Websockets.

Just to be complete I also mention ‘Long/Short polling’ as a client-side pull option. I regard it as mere workaround to a realtime communication implementation.

As always, requirements are key:

If you need bi-directional realtime features with high volume and low latency, Websocket is your choice.
If you need to deliver realtime information to many clients, but you can live with the HTTP Request/Response model to trigger actions, Server Sent Events (SSE) are a powerful contender.

SSE is a mechanism that allows the server to asynchronously push the data to the client once the client-server connection is established. The client subscribes to the server data via an Event Source.

Task Scheduling with limited Resources

Sun, 24 May 2020 00:00:00 +0000

We already looked at a simple scheduling problem: Task Scheduling.

Here we are looking at a more elaborated scheduling example with limited resources for tasks with various durations. We want to get insights by re-engineering the solution of ¹ with Pyomo.

Problem

We have $N$ tasks and $M$ facilities to execute the tasks.
Every task has a due date.
Tasks require certain resources for execution, e.g. water, electricity, …
Every facility provides a set of resources
Only one task per resource can be executed at a given time.

Assign the tasks to the facilities so that the task get all required resources for execution. We want to minimize the total time to complete all tasks. This time is called makespan.

Task Scheduling

Sat, 23 May 2020 00:00:00 +0000

We already looked at a real world scheduling problem: Patient Scheduling.

Important concepts have been touched there. Here we look at additional aspects of this important optimization domain, especially in using Pyomo as modelling environment.

Simple Scheduling

Problem

Schedule 10 tasks. Each task must be assigned to a single time slot, however, not every time slot is viable. The duration of a task is one timeslot. We want to minimize the number of timeslots used.

Polyominos, Tetris

Sun, 17 May 2020 00:00:00 +0000

Problem

Fill a rectangle with polyominos.

What is a Polyomino?
A plane geometric figure formed by joining one or more equal squares edge to edge. It is a polyform whose cells are squares. It may be regarded as a finite subset of the regular square tiling.

Model

Since this puzzle is just another form of tiling challenge we can reuse concepts and ideas from Mondriaan Puzzle, Another Boring Lockdown Day and Patient Scheduling.

Patient Scheduling

Sat, 09 May 2020 00:00:00 +0000

How to optimize the scheduling of patients in order to use the necessary infrastructure, i.e. infusion chairs optimally.

Problem Parameters ¹:

There are $T = 40$ time slots of 15 minutes
We have 23 infusion chairs for patient treatment
We have 8 different types of patients
Each patient type has a demand (number of patients) and treatment length (expressed in 15 minute slots)
No treatments can start during lunch break
At most 2 treatments can start in each time slot.

Yet Another Mathprogramming Consultant provides a comprehensive treatment of the problem. Here I will focus on a Mixed Integer Programming ‘grid approach’.

Square Tiling

Sat, 02 May 2020 00:00:00 +0000

Given an inventory of square tiles, what is the maximum square space we can fill with them?

The picture above consists of the following set of tiles:

length	1	2	3
count	4	3	2

8 out of the 9 given tiles can be used to form the resulting (5x5) square. One (3x3) tile cannot be used. With the given set this is the larges square area we can fill.

Mondriaan Puzzle

Fri, 01 May 2020 00:00:00 +0000

Partition an $n \times n$ square into multiple non-congruent integer-sided rectangles.
The rectangles must be non-congruent and the side length must be an integer number.
Find the least possible length difference between the largest and smallest used rectangle, i.e. make the rectangles as similar as possible.
Every rectangle can only be used once.

There is related work covering multiple aspects of the problem and its solution. Here we focus on a Mixed Integer Programming approach.

Sudoku

Sun, 26 Apr 2020 00:00:00 +0000

Wikipedia

a logic-based, combinatorial number-placement puzzle.

Fill a 9×9 grid with digits so that each column, each row, and each of the nine 3×3 subgrids that compose the grid contain all of the digits from 1 to 9.

The puzzle setter provides a partially completed grid, which for a well-posed puzzle has a single solution.

Challenges

Understand the problem (easy)
Design the index sets (medium)
Translate the model into Pyomo (easy)

Tooling

Pyomo as LP modelling language
Optimizer: CBC
Python

Model

Indexing Schema

Variables

$$ y^m_{i,j;k} = \begin{cases} 1, \ \text{ when cell $(i,j;k)$ contains number $m$ }\\ 0, \ \text{ else }\\ \end{cases} $$

Longest Path Journey

Sat, 25 Apr 2020 00:00:00 +0000

Normally you would be searching for the shortest path from A to B. At least this is what navigation systems do. But here we want to look at the opposite problem:

What is the longest possible path to get from A to B?

Wikipedia

The longest path problem is the problem of finding a simple path of maximum length in a given graph. In contrast to the shortest path problem, which can be solved in polynomial time in graphs without negative-weight cycles, the longest path problem is NP-hard.

Finding the way with TSP

Fri, 17 Apr 2020 00:00:00 +0000

Travelling Salesman (TSP) Art in Python

There is already a lot of existing work on this topic ¹, but especially the book from Robert Bosch ² inspired me to embark on my own journey.

It became an rewarding and entertaining melange of “Mathematical Optimization” and playful programming while exploring new concepts and building a basic theoretical foundation for one of the most famous problems in mathematical optimization.

The Travelling Salesman Problem (TSP)

Starting from his home, a salesman wishes to visit each of $(n − 1)$ cities and return home at minimal cost. He must visit each city exactly once and it costs $c_{ij}$ to travel from city $i$ to city $j$. What route should he select?

Optimized Car Rental

Sat, 11 Apr 2020 00:00:00 +0000

Modelling a car rental company with medium complexity or how to make 120k € per week.¹

If you need a primer on Linear Programming: Linear Programming for Dummies 1

Tools employed:

Linear programming
Python
Pyomo as LP modelling language
Optimizer: CBC

Company model

Let’s assume a car rental company with 4 locations and business hours from Monday to Saturday.

The rental car demand estimate is:

Cars can be rented for 1, 2 or 3 days and can be returned to either the depot from which rented or another depot at the start of the next morning. A 3-day rental on Friday means that the car has to be returned on Tuesday morning because Sunday is closed.

BMW I4 and linear dynamic Truchet tilings

Sat, 28 Mar 2020 00:00:00 +0000

Isn’t this BMW I4 a beauty? The perfect motivation to start a journey into a little optimization problem:

How to manipulate traditional Truchet tiles so that the resulting tiling resembles the Ultimate Driving Machine?

Truchet tiles are a well known niche in math because they form a foundation for a lot of interesting questions. The basic theory: You have four simple tiles:

Since every tile is 50% black and 50% white a resulting tiling can only resemble a grey area if viewed from afar. In order to depict images it is necessary to manipulate the proportion of black vs. white.

Munggoggo: A modern message based async agent framework

Sat, 02 Nov 2019 22:12:03 +0000

An asyncio based agent platform written in Python and based on RabbitMQ.

Agents are isolated processes which can host multiple parallel running behaviours implementing business logic and can be deployed as independent kubernetes pods. Access to the agent mesh is via REST/HTTP, websocket and jsonrpc.

A few lines of code implement a fully fledged agent with a simple behaviour: sending and receiving ping messages to/from other agents.

from mode import Worker
from behaviour import Behaviour
from core import Core


class Agent(Core):
    class PingBehav(Behaviour):
        async def setup(self):
            self.counter = 0

        async def run(self):
            self.counter += 1
            msg = await self.receive()
            if msg:
                print(f"{self.name}: Message received: {msg.body.decode()}")
            await self.publish(str(self.counter), 'ping')
            await asyncio.sleep(0.9)

    async def setup(self) -> None:
        """ Register behaviour and subscribe to 'ping' topic """
        await self.add_runtime_dependency(self.PingBehav(self, binding_keys=['ping']))


if __name__ == '__main__':
    Worker(Agent(identity='AgentIdentity'), loglevel="info").execute_from_commandline()

Documentation: https://munggoggo.readthedocs.io/en/latest/
Github: https://github.com/sysid/munggoggo

Energy Forecast for a full-scale Vehicle Plant

Sun, 28 May 2017 22:12:03 +0000

Energy Forecast for a full scale Vehicle Plant

Energy forecasting is based on time series analysis. There are many techniques for analysing and forecasting time series, e.g. ARIMA, linear regression and deep learning. To tackle the challenge at hand a linear regression will be the benchmark model aganst which deep learning models will be tested. In particular a multi layer perceptron (MLP) and recurrent neural network (RNN), i.e. Long-Short Time Memory (LSTM) model will be applied.

Fishy Affine Transformation

Mon, 13 Mar 2017 22:12:03 +0000

Fishy Affine Transformation

While working on the kaggle competition https://www.kaggle.com/c/the-nature-conservancy-fisheries-monitoring I hit the point when I wanted to align fish based on an annotation at the fish’s head and tail, so that the fish is centered in the image, always in the same orientation and distracting picture information is minimized. This required:

finding the fish (thanks Nathaniel Shimoni for annotating)
centering
rotatating
cropping

Mathematically the challenge is to find the associated Affine Transformation. After years of working in a managerial role my linear algebra skills are a bit rusty so I decided to invest the weekend.

Machine Learning Journey

Mon, 23 Jan 2017 22:12:03 +0000

Cheat Sheet

General Explanations:

embeddings: a way to translate multidimensional input into fixed length log dimensional representations: lookup the integer index of the object and look it up in a corresponding matrix wich holds the low-dim representation. If no embeddings are used, the input has to be one-hot-encoded wich yields huge matrices
KFold Cross Validation: The purpose of cross-validation is model checking, not model building. Once we have used cross-validation to select the better performing model, we train that model (whether it be the linear regression or the neural network) on all the data. We don’t use the actual model instances we trained during cross-validation for our final predictive model.
A dense layer in a multilayer perceptron (MLP) is a lot more feature intensive than a convolutional layer. People use convolutional nets with subsampling precisely because they get to aggressively prune the features they’re computing.
in NNs rarely occur local minima due to vast parameter space (probability not to get better in ayn dimension is miniscule)
the fast majority of space of a loss function in NN is all saddlepoints
one training cycle for the entire dataset is called epoch, i.e. the algorithm sees the ENTIRE dataset
iteration: every time a batch is passed through the NN (forward + backward pass)
Latent factors = features of embeddings (used in Collaborative Filtering)
Bias-Variance Tradeoff in Machine Learning
Softmax vs Sigmoid: All the softmax units in a layer are constrained to add up to 1, whereas sigmoid units don’t have this ’lateral’ constraint. If every example can be associated with multiple labels, you need to use a sigmoid output layer that learns to predict “yes” or “no” for each individual label. If the classes are disjoint, i.e. each example can only belong to one class, you should use a softmax output layer to incorporate this constraint.
Do not forget to fine tune your network architecture and your learning rate. If you have more data, a complex network is preferable. According to one important deep learning theorem, the local minima are very close to the global minimum for very deep neural networks.

Preprocessing

pre-process your data by making sure each dimension has 0 mean and unit variance. This should always be the case with data your are feeding to a NN, unless you have strong, well-understood reasons not to do it. A simple MLP will never cause gradient explosion if your data is correctly preprocessed.
Centering sparse data would destroy the sparseness structure in the data, and thus rarely is a sensible thing to do.
However, it can make sense to scale sparse inputs, especially if features are on different scales.
MaxAbsScaler and maxabs_scale were specifically designed for scaling sparse data, and are the recommended way to go

Underfitting/Overfitting

Underfitting: This describes a model that lacks the complexity to accurately capture the complexity inherent in the problem you’re trying to solve. We can recognize this when our training error is much lower than our validation error
Overfitting: This describes a model that is using too many parameters and has been trained too long. Specifically, it has learned how to match your exact training images to classes, but has become so specific that it is unable to generalize to similar images. This is easily recognizable when your training set accuracy is much higher than your validation.
when you start overfitting you know, that your model is complex enough to handle your data
Your main focus for fighting overfitting should be the entropic capacity of your model –how much information your model is allowed to store. A model that can store a lot of information has the potential to be more accurate by leveraging more features, but it is also more at risk to start storing irrelevant features. Meanwhile, a model that can only store a few features will have to focus on the most significant features found in the data, and these are more likely to be truly relevant and to generalize better.
Dropout also helps reduce overfitting, by preventing a layer from seeing twice the exact same pattern, thus acting in a way analoguous to data augmentation (you could say that both dropout and data augmentation tend to disrupt random correlations occuring in your data).

Recipe:

Python on sysid blog

How Claude Code Saved My Weekend

Modern Open-Source Maintenance

Anki, Markdown, and the Power of inka2

Streamlining Knowledge Management: Anki, Markdown, and the Power of inka2

Introduction

Why Markdown for Knowledge Management?

Linear Optimization in time-critical Control Loop

Problem

Model

Select Points

Model

Non-convex MIQP Model

Linear MIP Model

Software Architecture 2022

CRUD, MVC

DDD and Clean Architecture

Maximal Overlap of Rectangles

Problem Parameters:

Model

Parameters

Scheduling

Problem 1: Machine Scheduling

Inventory Management

Cost Contribution

Job Shop Scheduling Problem

Resource Constrained Project Success Plan

Problem Domain: Network Optimization.

Installing CBC on MacOS

Update May 2022: Using the coinbrew approach as documented here works now out of the box.

Dispersion

Model

The Book of Mormon II

Recap Challenge

Objective:

The Book of Mormon

Challenge

String Puppets

Numberlink Puzzle

Problem

Solution

Server Sent Events

As always, requirements are key:

Task Scheduling with limited Resources

Problem

Task Scheduling

Simple Scheduling

Problem

Polyominos, Tetris

Problem

Model

Patient Scheduling

Problem Parameters 1:

Square Tiling

Mondriaan Puzzle

Sudoku

Challenges

Tooling

Model

Indexing Schema

Variables

Longest Path Journey

What is the longest possible path to get from A to B?

Finding the way with TSP

Travelling Salesman (TSP) Art in Python

The Travelling Salesman Problem (TSP)

Optimized Car Rental

Tools employed:

Company model

BMW I4 and linear dynamic Truchet tilings

Munggoggo: A modern message based async agent framework

Energy Forecast for a full-scale Vehicle Plant

Energy Forecast for a full scale Vehicle Plant

Fishy Affine Transformation

Fishy Affine Transformation

Machine Learning Journey

Cheat Sheet

General Explanations:

Preprocessing

Underfitting/Overfitting

Problem Parameters ¹: