Rubie Gauthier

Kepler just unveiled K2 "Bumblebee" at IROS 2025 in Hangzhou-and opened the stack to developers. This move isn't about another stage demo. It's a bid to win on ecosystem: exposing core OS elements, standardizing I/O, and pairing an app/toolbox with simulation and a future skills marketplace. Hardware matters, too-hybrid actuation with planetary roller screws prioritizes durability and smooth, humanlike joint motion. Context worth watching: IROS 2025 (Oct 19–25, Hangzhou) underscored how fast humanoid R&D is globalizing. And sim-to-real is delivering usable throughput. In November 2024, Fourier trained a floor-to-stand policy in ~3,000 iterations (~15 hours) using NVIDIA Isaac Gym, then transferred it with 89% success-evidence that teams can compress iteration cycles without burning hardware time. Compute headroom is also shifting. Jetson AGX Thor offers up to 7.5× the AI compute of its predecessor, making on-robot inference for richer perception and whole-body control more viable for third-party devs. If Kepler's developer program lands with clear APIs, sandboxed safety, and fair revenue share, we'll finally see reusable "skills" emerge beyond research labs.

3

Great piece by Emilia David on the LangChain open source integration ecosystem, the power of LangGraph, and our newly released LangGraph Platform Helps put everything you see us working in perspective Link: https://lnkd.in/ghajS7zx

100

5 Comments

A different path to business succession. Not family.　Not employees.　Not M&A. Releasing 30 years of software as open source and sustaining the business through a community. Press release https://lnkd.in/gD3-GWTN #OpenSource #Manufacturing #BusinessSuccession

1

Attended the first ever FoundersBeta event this week. Panel with Viler Lika, MBA, B.Eng Lika (Founder & CEO, SingleKey), Mark Bania (Founder and exited former CEO, Contractor Compliance), and John Marzo (Co-Founder & CEO, Airfairness). All bootstrapped founders. Refreshing perspective from people who didn't start with angel or VC money in their pocket. The panel focused on what actually works when you can't just spend on advertising, which a lot of "best practices" assume you can. Main theme was figuring out your ICP (Ideal Customer Profile), knowing where they live, and finding scrappy ways to get their attention. One example stuck with me. Automated social listening for pain points then responding in real time, showing how your product solves their specific problem while they still feel the frustration. Simple but effective. Thanks to Eric Rafat and the FoundersBeta team for putting this together.

52

3 Comments

I'm incredibly excited for this talk. Top K queries are the backbone of search engines, which have historically been NoSQL databases unable to execute JOINs. We're fixing this with ParadeDB. Highly recommend attending, I promise you'll learn something!

36

A bit later than the official launch of Flower Hub from Flower Labs. Many thanks to Yan Gao for inviting me to participate as an early contributor. It was the push I needed to dust off an old pet project I'd been putting aside for way too long: federated UMAP. Before you train a model, you need to understand your data. Clustering and dimensionality reduction help with that enormously. But for UMAP you need pairwise distances between all data points — which means centralizing everything. That pretty much kills the idea of federated learning, where you don't share data. Fair point. But one can use approximations. The approach (from Qiao, Ma & Fan, AAAI 2025): clients collaboratively learn a shared set of landmark points via MMD optimization. The server reconstructs the global distance matrix using Nyström approximation and runs UMAP on the result. No raw data ever leaves any client and the embeddings are close enough to centralized UMAP. The beauty is that this Nyström-based distance reconstruction is algorithm-agnostic - it would work with t-SNE, spectral clustering, or any method that relies on a pairwise distance matrix. I chose UMAP specifically because it's widely used in biomedical research. For example, in single-cell RNA sequencing, where UMAP helps researchers identify cell types and disease subtypes across patient cohorts that often can't be pooled due to privacy regulations. The app works out of the box and logs embeddings directly to W&B for easy comparison. Huge thanks to the authors for their research, it's awesome: https://lnkd.in/duzMCG65 Congratulations Flower Labs on their launch, future is federated :) App itself can be found here: https://lnkd.in/d45YzPeX

14

If your audience can't do something useful with your product in less than 10 mins, you are at a significant disadvantage. No amount of DevRel will fix this.

4

🦀 Proud to share — Undercurrent News featured SeafoodAI’s launch of CrabScan360, focusing on how our AI biometric seafood scanner empowers better decisions and compliance for crabs. Manual methods are slow, subjective, and expensive. CrabScan360 changes that. It digitizes the process from dock to plant. Here’s what we’re delivering: ✅ Instant crab size and weight detection ✅ Automated compliance checks ✅ Real-time data for certification and inventory management   It’s all part of our push to modernize one of the world’s most under-digitized supply chains. We’re starting with crabs, and expanding from there. 📖 Read the full article: https://lnkd.in/gfpyp3ub Big thanks to our early pilot partners helping bring this vision to life. #SeafoodAI #BlueEconomy #AI #SeafoodTech #SustainableSeafood #ClimateTech #Traceability #DigitalTransformation #NECX #Undercurrentnews

18

4 Comments

At a recent AI convening, two things that struck me had almost nothing to do with AI. Early on, Murugan Vasudevan from Veddis Foundation spoke about Kunji, a Glific-based WhatsApp bot used by government officials to access policy-related information. In a Q&A following his talk, it became clear that for this approach to work (or be scaled up) one first has to streamline and structure all the policy-related documentation, which could be spread across multiple departments. This consolidated set of documents is then used to train the bot. So building the bot is the easy part – the hard bit is the cleanup needed upfront, which also includes setting up processes to keep this document repository updated. All this is just good practice, which can yield benefits even without AI. And without this preparation, even the best intelligence is of little use. Continue reading... Project Tech4Dev Veddis Foundation Give Atlassian #genai #ai https://lnkd.in/g5WX2K2M

20