Date:
Venue:	Unknown
City:
Time:
Ages:
Opener:
Tickets:
Notes:	Unknown
Other Links:	Tell a friend!

A few weeks ago I shipped NewsBlur v14 for iOS and Mac, a major redesign of the Apple apps. Today, Android gets the same treatment. Every screen has been reworked: the feed list, the story list, the reading view, preferences, and menus. Along with the visual overhaul, several features that were previously web-only are now on Android: Ask AI, Discover Related Sites, and the Daily Briefing.

Here’s what’s new.

Ask AI

Ask AI brings the same AI-powered Q&A from the web and iOS to Android. Open any story, tap Ask AI, and ask questions about it. Summarize a long article, get background on a developing situation, or fact-check a claim. Pick your preferred AI model and keep the conversation going with follow-ups. The Ask AI sheet matches your current theme and slides up as a bottom sheet, consistent with the share and trainer dialogs.

Discover related sites

Discover Related Sites lets you find new feeds related to any feed you’re already subscribed to. Tap the Discover button in the story list header bar, browse what’s available, and preview a feed before subscribing. Duplicate feeds are filtered out so you only see new options.

Daily Briefing

The Daily Briefing generates a personalized summary of your news, organized into sections like Top Stories, Based on Your Interests, and Long Reads. It uses native Android story rows, so it feels like a regular feed rather than a bolted-on feature. Configure your briefing frequency, writing style, and sections from the briefing view in your sidebar.

Sepia theme and refined dark themes

A new Sepia theme brings warmer tones for comfortable long reading sessions. The Dark theme has been lightened to match the iOS gray/medium palette, and the Black theme now uses true absolute black backgrounds for feed and story cells, making it ideal for OLED screens.

Story list header bar

The top of the story list now has a header bar with quick access to Discover, search, display options, and settings. The display and settings controls are split into separate menus so you can change the view without wading through unrelated options.

Redesigned reading experience

The reading view has been rethought from top to bottom. Story traversal buttons are lifted above the bottom edge for easier thumb access. A new traverse bar with refined icons shows your position and unread count. Story actions are hidden until the story finishes rendering, so you never tap a button before the content is ready. Opening a story from the list now animates smoothly into the reader, and swiping back uses an interactive gesture that tracks your finger.

Redesigned preferences and menus

Preferences have been rebuilt as a modern settings screen with inline segments instead of separate dialog pickers. The feed list menu, reading menu, and folder menus have all been redesigned with cleaner styling and better organization. Menus now scale with your device font size, so they stay readable at any accessibility setting.

Premium Archive and Pro subscriptions

You can now subscribe to Premium Archive and Premium Pro directly from the Android app. An upgrade banner appears in the story list when you’re on a lower tier, showing what you’d unlock by upgrading.

Everything else

Beyond the headline features, this release includes a long list of improvements and fixes.

Improvements

• Interactive swipe-back gesture in both the story list and reading view with predictive back support on Android 14+.
• Feed list aligned with iOS styling, with new collapse-all and expand-all toggles.
• Story header pills with compact layout and title case formatting.
• Active reading time tracking per story, synced to your account.
• Full text and regex classifiers for the Intelligence Trainer.
• Feed search field themed to match your current theme with autofill disabled.
• Sync done pill delayed until feeds actually render, so you see the update happen.
• Story thumbnails enlarged for small sizes and cropping fixed.
• Status banners at the top of the story list for loading and error states.
• Mute Sites redesigned with upgrade card and progress bar.
• Custom folder and feed icon support.

Fixes

• Fixed TransactionTooLargeException crash in the reading pager.
• Fixed database version mismatch crash on launch.
• Fixed ItemListMenuPopup crash on small and split-screen displays.
• Fixed login autofill and app switching losing input.
• Fixed story row thumbnail cropping.
• Fixed search pill text vanishing.
• Fixed story list edge back gesture interference.
• Brightened story feed titles for better readability.

Coming up next: v14.2 will bring story clustering to Android, so duplicate stories across your feeds get grouped together automatically, just like on the web.

NewsBlur v14 for Android is available now on the Google Play Store. If you have feedback or run into issues, I’d love to hear about it on the NewsBlur forum.

Make sure your test files in a pytest tested project sit in directories with a __init__.py

Nothing ever made is truly perfect and indeed, CPU architectures like x86, RISC-V, ARM, and PowerPC all have their own upsides and downsides. Today, I aim to make an architecture that learns from all these mistakes and improves architecture design for everyone.

I’ve consulted with many people opinionated on the matter, both from a software perspective, and from a hardware perspective. I have taken all their feedback in mind while creating this initial draft of the WheatForce architecture (PDF). It is inspired by pieces from many architectures: segmentation inspired by x86, hash table-like paging from PowerPC, dynamic endianness control from RISC-V and PowerPC, and more. Let’s look into each feature in a little bit more detail.

Segmentation

Segmentation is a powerful virtual-memory feature that is tragically underused today. I believe this is due to limited flexibility, so I have added an improvement above the model that x86 had used: every single register can now use its own segment selector. With this added flexibility, one can surely make better use of the address translation powers of segmentation with minimal extra overhead.

Hash Table-Like Paging

PowerPC’s hash table-like paging makes its paging vastly superior to the likes of x86, RISC-V and ARM by decreasing the number of required cache line fetches drastically. Much like a true hash table, the keys (or input addresses) are hashed and then used as an index into the table. From there, that row of the table is searched for a cell with a matching virtual address, which can be accelerated greatly due to superior cache locality of the entries in this row.

Dynamic Endianness Control

RISC-V and PowerPC both have some real potential for better compatibility with their dynamic endianness control. However, both these architectures can only change the endiannes from a privileged context. To make this more flexible, WheatForce can change the data endianness at any time with a simple instruction. Now, user software can directly interoperate between big-endian and little-endian data structures, eliminating the need for a costly byte-swap sequence that would need many instructions. Finally, you can have your cake and eat it to!

Conclusion

WheatForce has observed the mistakes of all architectures before it, and integrates parts of all its predecessors. You can read the full specification on GitHub. After you’ve read it, do let me know what you think of it.

Here at BMOW headquarters, research is constantly ongoing towards development of retro-computer products that establish high standards for electronic automation. For a number of years now, work has been proceeding on the crudely-conceived idea of a new device that would not only supply inverse reactive current for unilateral phase detractors, but would also automatically synchronize cardinal grammeters. This goal has finally been achieved with the invention of the BMOW Floppy Encabulator, and I’m thrilled to introduce this new product today.

Conceptual Overview of Encabulation

Interest in encabulation technology has been growing steadily, but the underlying concepts may be unfamiliar to some readers. Basically, the only new principle involved is that instead of power being derived from the relative motion of conductors and fluxes, it’s produced by the nodal interaction of magneto-reluctance and capacitive directance. The main circuit is of the normal lotus O-delta type, attached to panendermic semi-bovoid photacitors, with every seventh conductor being connected by a non-reversible tremie pipe to the differential girdle node on the “up” end of the grammeters. 

The operating point is maintained as near as possible to the H.F. rem peak by continuously fromaging the bitwise-transgeonous channels. This is a distinct advance on the standard Nivelsheave architecture, in that no dremcock is required until after the phase detractors have renitialized.

New Case Design Improves Stability

The device has a baseplate of pre-fabulated amulite, surmounted by a malleable logarithmic casing in such a way that the two main spurving regulators are in a direct line with the pentametric fan. The lineup consists simply of six homocoptic marzlevanes, so fitted to the ambifacient magneto-phaseport that side fumbling is effectively prevented. 

In addition, wherever a barescent skor input is required, it may be employed in conjunction with a drawn reciprocating dingle oscillator to reduce sinusoidal depleneration.

Performance Analysis and Relative Periodicosity

The 41 manestically-spaced grouting circuits are arranged to feed into the semioctal data stream with a superposition of high S-value elliptarithmic sequences and 5% ruminative impulsatrons. Both of these signals have specific periodicosities given by 

        P = 2*5 Cn(6*7) 

where n is the diahelical eigenphase of retrograde dislocation and C is Cholmondeley’s fundamental grillage coefficient. Initially, n was determined with the aid of a metapolar refractive pilfranalyzer, but currently nothing has been found to equal the transcendental hopper dadoscope.

Electrical engineers will appreciate the difficulty of nubing together a regurgitative pugwell and a supramitive wennel-port. Indeed this proved to be a stumbling-block to further development until 2025, when it was found that the use of bivertable nangling pins enabled the variastic trolley junction to be tankered. 

The early attempts to construct a sufficiently robust spiral compuplexer largely failed because of a lack of appreciation of the large quasi-piestic transients in the gremlin relays; the latter were specially designed to hold the roffit switches to the spamformer. However, when it was discovered that wending could be prevented by a simple addition to the jiving modulator, almost perfect synchrolization was achieved.

Coming Soon

The BMOW Floppy Encabulator has now reached a high level of technical development, and has already been successfully used for operating milford trenions. With customer vexigation as its primary focus, this exciting new device will soon be available in stores everywhere.

Each year at PyCon US, Startup Row highlights a select group of early-stage companies building ambitious products with Python at their core. The 2026 cohort reflects a rapidly evolving landscape, where advances in AI, data infrastructure, and developer tooling are reshaping how software is built, deployed, and secured.

This year’s companies aim to solve an evolving set of problems facing independent developers and large-scale organizations alike: securing AI-driven applications, managing multimodal data, orchestrating autonomous agents, automating complex workflows, and extracting insight from increasingly unstructured information. Across these domains, Python continues to serve as a unifying layer: encouraging experimentation, enabling systems built to scale, and connecting open-source innovation with real-world impact.

Startup Row brings these emerging teams into direct conversation with the Python community at PyCon US. Throughout the conference, attendees can meet founders, explore new tools, and see firsthand how these companies are applying Python to solve meaningful problems. For the startups in attendance, it’s an opportunity to share their work, connect with users and collaborators, and contribute back to the ecosystem that helped shape them. Register now to experience Startup Row and much more at PyCon US 2026.

CapiscIO

As multi‑agent AI systems become the backbone of emerging digital workflows, developers lack a reliable way to verify agent identities and enforce governance. CapiscIO steps into that gap, offering an open‑core trust layer built for the nascent agent economy.

CapiscIO offers cryptographic Trust Badges, policy enforcement, and tamper‑evident chain‑of‑custody wrapped in a Python SDK. Released under Apache 2.0, it ships a CLI, LangChain integration, and an MCP SDK that let agents prove identity without overhauling existing infrastructure.

The capiscio‑core repository on GitHub hosts the open‑source core and SDKs under Apache 2.0, drawing early contributors building agentic pipelines.

Beon de Nood, Founder & CEO, brings two decades of enterprise development experience and a prior successful startup to the table. “AI governance should be practical, not bureaucratic. Organizations need visibility into what they have, confidence in what they deploy, and control over how agents behave in production,” he says.

CapiscIO is continuously adding new extensions, expanding its LangChain and MCP SDKs, and preparing a managed agent‑identity registry for enterprises. In short, CapiscIO aims to be the passport office of the agent economy, handing each autonomous component an unspoofable ID and clear permissions.

Chonkie

The explosion of retrieval‑augmented generation (RAG) is unlocking AI’s ability to reason over ever‑larger knowledge bases. Yet the first step of splitting massive texts into meaningful pieces still lags behind.

Chonkie offers an open‑core suite centered on Memchunk, a Python library with Cython acceleration that delivers up to 160 GB/s throughput and ten chunking strategies under a permissive license. It also ships Catsu, a unified embeddings client for nine providers, and a lightweight ingestion layer; the commercial Chonkie Labs service combines them into a SaaS that monitors the web and synthesises insights.

Co‑founder and CEO Shreyash  Nigam, who grew up in India and met his business partner in eighth grade, reflects the team’s open‑source ethos, saying “It’s fun to put a project on GitHub and see a community of developers crowd around it.” That enthusiasm underpins Chonkie’s decision to release its core tooling openly while building a commercial deep‑research service.

Backed by Y  Combinator’s Summer  2025 batch, Chonkie plans to grow from four to six engineers and launch the next version of Chonkie Labs later this year, adding real‑time web crawling and multi‑modal summarization. In short, Chonkie aims to be the Google of corporate intelligence.

Pixeltable

Multimodal generative AI is turning simple datasets into sprawling collections of video, images, audio and text, forcing engineers to stitch together ad‑hoc pipelines just to keep data flowing. That complexity has created a new bottleneck for teams trying to move from prototype to production.

The open‑source Python library from Pixeltable offers a declarative table API that lets developers store, query and version multimodal assets side by side while embedding custom Python functions. Built with incremental update capabilities, combined lineage and schema tracking, and a development‑to‑production mirror, the platform also provides orchestration capabilities that keep pipelines reproducible without rewriting code.

The project has earned ≈1.6 k  GitHub stars and a growing contributor base, closed a $5.5  million seed round in December 2024, and is already used by early adopters such as Obvio and Variata to streamline computer‑vision workflows.

Co‑founder and CTO Marcel  Kornacker, who previously founded Apache Impala and co-founded Apache Parquet, says “Just as relational databases revolutionized web development, Pixeltable is transforming AI application development.”

The company's roadmap centers on launching Pixeltable Cloud, a serverless managed service that will extend the open core with collaborative editing, auto‑scaling storage and built‑in monitoring. In short, Pixeltable aims to be the relational database of multimodal AI data.

Skyvern

Manual browser work remains a hidden bottleneck for many teams, turning simple data‑entry tasks into fragile scripts that break on the slightest UI change. Skyvern’s open‑source agent is one of the tools reshaping how developers and non‑technical users automate the web.

The Skyvern library lets anyone build a no‑code browser agent that combines computer‑vision models with a large language model to see, plan, act, and validate each step of a web workflow. Its planner–actor–validator loop compiles successful runs into deterministic code, while the free open‑source core can be run locally or via Skyvern Cloud on a per‑automation pricing model.

The GitHub repository has attracted ≈20 k stars, drawing an active community of contributors who extend the framework and share evaluation datasets. The company monetizes through Skyvern Cloud, letting teams run agents without managing infrastructure.

Skyvern is preparing a release that tightens vision‑model integration, adds support for additional LLM providers, and launches a self‑serve dashboard aimed at non‑technical teams. In short, Skyvern aspires to be the Django of browser‑automation, pairing developer friendliness with production reliability.

SubImage

The sheer complexity of modern multi‑cloud environments turns security visibility into a labyrinth, and  SubImage offers a graph‑first view that cuts through the noise.

It builds an infrastructure graph using the open‑source Cartography library (Apache‑2.0, Python), then highlights exploit chains as attack paths and applies AI models to prioritize findings based on ownership and contextual risk.

Cartography, originally developed at Lyft and now a Cloud Native Computing Sandbox project, has ≈3.7 k GitHub stars, is used by over 70 organizations, and SubImage’s managed service already protects security teams at Veriff and Neo4j; the company closed a $4.2  million seed round in November 2025.

Co‑founder Alex  Chantavy, an offensive‑security engineer, says “The most important tool was our internal cloud knowledge graph because it showed us a map of the easiest attack paths … One of the most effective ways to defend an environment is to see it the same way an attacker would.”

The startup is focusing on scaling its managed service and deepening AI integration as it targets larger enterprise customers. In short, SubImage aims to be the map of the cloud for defenders.

Tetrix

Private‑market data pipelines still rely on manual downloads and spreadsheet gymnastics, leaving analysts chasing yesterday’s numbers. Tetrix’s AI investment intelligence platform is part of a wave that brings automation to this lagging workflow.

Built primarily in Python, Tetrix automates document collection from fund portals and other sources, extracts structured data from PDFs and other unstructured sources using tool-using language models, then presents exposures, cash flows, and benchmarks through an interactive dashboard that also accepts natural‑language queries.

The company is growing quickly, doubling revenue quarter over quarter and, at least so far, maintains an impressive record of zero customer churn. In the coming year or so, Tetrix plans to triple its headcount from fifteen to forty‑five employees.

TimeCopilot

Time‑series forecasting has long been a tangled mix of scripts, dashboards, and domain expertise, and the recent surge in autonomous agents is finally giving it a unified voice. Enter TimeCopilot, an open‑source framework that brings agentic reasoning to the heart of forecasting.

The platform, built in Python under a permissive open‑source license, lets users request forecasts in plain English. It automatically orchestrates more than thirty models from seven families, including Chronos and TimesFM, while weaving large language model reasoning into each prediction. Its declarative API was born from co‑founder  Azul Garza‑Ramírez’s economics background and her earlier work on TimeGPT for Nixtla (featured SR'23), evolving from a weekend experiment started nearly seven years ago.

The TimeCopilot/timecopilot repository has amassed roughly 420  stars on GitHub, with the release of OpenClaw marking a notable spike in community interest.

Upcoming plans include a managed SaaS offering with enterprise‑grade scaling and support, the rollout of a benchmarking suite to measure agentic forecast quality, and targeted use cases such as predicting cloud‑compute expenses for AI workloads.

Thank You's and Acknowledgements

Startup Row is a volunteer-driven program, co-led by Jason D. Rowley and Shea Tate-Di Donna (SR'15; Zana, acquired Startups.com), in collaboration with the PyCon US organizing team. Thanks to everyone who makes PyCon US possible.

We also extend a gracious thank-you to all startup founders who submitted applications to Startup Row at PyCon US this year. Thanks again for taking the time to share what you're building. We hope to help out in whatever way we can.

Good luck to everyone, and see you in Long Beach, CA!