Python Hub Weekly Digest for 2025-09-28
This week in Python, popular projects included "marker" for converting PDF to markdown, "Nvmath-Python" offering Nvidia Math Libraries for Python, and "riffq", a toolkit for building PostgreSQL databases in Python. Interesting articles covered topics such as Python signals/MIDI processing system "Nallely", Microsoft Python Driver for SQL Server, and a tutorial on building an AI-assisted Reddit scraping pipeline. Additionally, PyCon Australia 2025 talks videos are now available. In the projects section, "TensorRT-Model-Optimizer", "Tiny LLM", and "detroit", a Python implementation of d3js, were highlighted. Have a great week and happy coding!
đź’– Most Popular
datalab-to / marker
Convert PDF to markdown + JSON quickly with high accuracy
List of 87 Programming Ideas for Beginners (with Python implementations)
Nvmath-Python: Nvidia Math Libraries for the Python Ecosystem
Jaxformer Scaling Modern Transformers
This is a zero-to-one guide on scaling modern transformers with n-dimensional parallelism. Transformers have driven much of the deep learning revolution, yet no practical guide reflects SOTA architectures and the complexities of large-scale language modelling. While excellent resources such as DeepMind’s How to Scale Your Model and HuggingFace’s Ultra Scale Playbook exist, a gap remains ...
riffq
A toolkit for building PostgreSQL wire-compatible databases in Python, powered by Rust for performance and concurrency.
đź“– Articles
Nallely – A Python signals/MIDI processing system inspired by Smalltalk
Microsoft Python Driver for SQL Server
PyCon Australia 2025
PyCon Australia 2025 talks videos are available now.
Sphinx Docs Instantly in Your Browser (MyST Markdown + reStructuredText)
Edit and preview reStructuredText or MyST Markdown instantly in a Sphinx running in a browser. Runs entirely in Python using WebAssembly, so it’s private, fast, and ideal for learning markup.
Python Tutorial: Build an AI-assisted Reddit Scraping Pipeline
The video provides an in-depth, hands-on tutorial for building a resilient, AI-assisted Reddit scraping pipeline in Python, covering everything from Jupyter prototyping and LangChain agents to a Django-based background worker architecture. It teaches viewers to automate web scraping, integrate Google’s Gemini LLM for query refinement, and store structured results in PostgreSQL, suitable ...
Post-training 101
A hitchhiker's guide into LLM post-training.
Avoid Messy Code: Design Patterns for AI Agents in Python
The video demonstrates how to keep Python code for AI agents clean and maintainable by applying design patterns like Chain of Responsibility (for modular pipelines), Observer (for agent logging and context), and Strategy (for pluggable agent personalities). These patterns help decompose logic, improve scalability, and ensure testability for complex AI workflows.
Just for fun: animating a mosaic of 90s GIFs
The post describes an experiment in animating a mosaic of vintage 90s GIFs collected from the GeoCities archive, using HTML Canvas for random, lively playback. It celebrates the playful aesthetics of early web graphics and highlights the technical and nostalgic joy of reintroducing these classic GIFs into a modern browser setting.
Hyperparameter Tuning Tips that 99% of Data Scientists Overlook
This video shows how to tune XGBoost models with Optuna while maximizing speed using XGBoost 3.0’s GPU acceleration for 5–15x faster training. He explains why cross-validation is crucial, recommends smart tuning practices, and demonstrates how Optuna’s visualizations help identify impactful hyperparameters in real-world tabular data workflows.
Defeating Nondeterminism in LLM Inference
LLM inference is often nondeterministic even with temperature set to zero, primarily due to batch-size-dependent kernel behaviors that change results based on server load rather than randomness or floating-point issues. The solution is to use batch-invariant kernels, ensuring reproducible outputs even in high-concurrency environments, which is now possible but may come with some efficien...
Context Engineering - Short-Term Memory Management with Sessions from OpenAI Agents SDK
The guide demonstrates how to use the OpenAI Agents SDK’s Session object to manage short-term memory in AI agents, enabling context trimming and compression for efficient, coherent, and cost-effective multi-turn conversations. Effective session memory ensures agents maintain relevant history across turns while reducing noise, latency, and error risk in longer interactions.
Tricks from OpenAI gpt-oss YOU ?? can use with transformers
The post details major upgrades that allow models like OpenAI’s GPT-OSS to run, fine-tune, and scale efficiently, including zero-build kernels, 4-bit MXFP4 quantization, tensor and expert parallelism, dynamic layerwise caching, and continuous batching. These improvements cut memory usage, boost speed, and enable larger models to run on affordable hardware, making cutting-edge techniques ...
Alibaba-NLP / DeepResearch
Tongyi Deep Research, the Leading Open-source Deep Research Agent
Python Hub Weekly Digest for 2025-09-21
⚙️ Projects
TensorRT-Model-Optimizer
A unified library of state-of-the-art model optimization techniques like quantization, pruning, distillation, speculative decoding, etc. It compresses deep learning models for downstream deployment frameworks like TensorRT-LLM or TensorRT to optimize inference speed.
Tiny LLM - LLM Serving in a Week
A course of learning LLM inference serving on Apple Silicon for systems engineers: build a tiny vLLM + Qwen.
detroit
detroit is a Python implementation of d3js.
numethods
A lightweight, from-scratch, object-oriented Python package implementing classic numerical methods.
VeritasGraph
Enterprise-Grade Graph RAG for Secure, On-Premise AI with Verifiable Attribution.
Semlib
Build data processing and data analysis pipelines that leverage the power of LLMs.
ApeRAG
Production-ready GraphRAG with multi-modal indexing, AI agents, MCP support, and scalable K8s deployment
MathFlow
Likerequestsfor mathematical computing, making complex math feel simple.
AuthTuna
The Modern Async Security Framework for FastAPI.
JiraTUI
A Textual User Interface for interacting with Atlassian Jira from your shell.
ROMA
A meta-agent framework to build high-performance multi-agent systems.
Mini-o3
Scaling Up Reasoning Patterns and Interaction Turns for Visual Search.
đź‘ľ Reddits
Today I learned that Python doesn't care about how many spaces you indent as long as it's consistent
Python 3.13 is 10% slower than 3.12 for my file parser
Project by Ruslan Keba. Since 2012. Powered by Python. Made in 🇺🇦Ukraine.