Orange Collective

Honored to receive the Del Favero Thesis Prize at the Massachusetts Institute of Technology for the most innovative advance in the field of Nuclear Engineering ⚛️ My deepest gratitude goes to Professor Koroush Shirvan for his exceptional mentorship. I would also like to thank John Parsons and Jacopo Buongiorno for serving on my thesis committee, as well as all the NSE staff and faculty who made my time at MIT one of the most personally enriching and intellectually stimulating periods of my life. I would also like to express my appreciation to Benoit Forget and James Del Favero for this recognition. Now, it's time to build 🚀

Over the last two months, we've talked to more than 400 companies running agents in production. What's been most surprising: across all those industries, the agents that succeed fit into the same, three-part taxonomy. On Tuesday, we'll break down all of it. We'll cover customer-facing agent features, internal productivity tools, and developer platforms, along with the engineering challenges behind each. We'll also share what it takes to get an agent from demo to production. If you’d like to go even more in-depth, we've written two books on building AI systems. Principles explains the foundations of agents. Patterns covers how to ship them. Comment BOOK and we'll send you a copy of each. Comment BOX if you want copies for your team, class, hackathon, or event.

We’ve raised $5.2M to build the most accurate agentic tax prep system on the market. Most tax “automation” tools only solve the easy parts, while failing on or erroring on the hardest parts, meaning the firm still has to review every number in detail and can't actually save time, money, or stress during busy season. By combining our expertise in tax and technology, Magnetic has already reduced our customers’ time spent on tax returns by 77% - and we’re just getting started. This wouldn’t have been possible without all of our early customers, partners, and supporters - thank you for helping us get here. Special thanks to Formus Capital, Orange Collective, Seguin Ventures, Amino Capital, Singularity Venture Partners, Y Combinator and the dozens of angels including Jeff Seibert (CEO, Digits) and Kaz Nejatian (CEO, Opendoor) who participated in the round.

Starcloud 🤝 SpaceX

Inference engines today were designed primarily for text workloads. Multimodal support was added on top, with latency as an afterthought. We profiled Qwen3.6-27B on vLLM and found that video preprocessing takes ~500ms before inference ever ran. Under 5 QPS, this number grows to 11 seconds!! In this blogpost, Younes El Hjouji show how we cleared this bottleneck for both Qwen and Gemma models.

Congrats Luel on the $31.2M seed round led by General Catalyst and Lightspeed!