Romain Huet

Excited to announce that Pipeshift (YC S24) has raised a $2.5M seed round. Over the past year, enterprises have increasingly adopted open-source models for better privacy, control, and performance. New releases like DeepSeek R1 are driving even greater adoption, but deploying these models efficiently and securely remains a challenge. While models continue to improve, the tooling needed for engineering teams to train, deploy, and scale them in production is severely lacking. The amount of time and effort required to piece together multiple tools is significant, not to mention the effort needed to build new workflows. At Pipeshift, we're building a modular orchestration platform that allows teams to build with open source AI components, including LLMs, audio models, image models, vision models, embeddings, and vector DBs. Teams can deploy these models on their infrastructure, whether cloud or on-prem, while optimizing for performance, cost, and quality. In recent months, we've helped several teams train and deploy use case specific models for their workflows, and we’ll continue to do so. If you're currently building with open source models or thinking about making the switch, let's chat. We’re incredibly grateful to our earliest supporters at Y Combinator (GPs: Dalton Caldwell, Tyler Bosmeny, Michael Seibel, David Lieb), SenseAI Ventures, Arka Venture Labs, Good News Ventures, Nivesha Ventures, Astir Ventures, MyAsiaVC, gradCapital, 100X.VC and amazing angels including Kulveer Taggar, Umur Cubukcu, Krishna M., Nate Matherson, Max Kolysh, Fred Zhang, Ruchir Patwa, Arjun Maheswaran, Kishore Ganji, Praneeth Kannah, Anshul Bhide, Hasan Abu Sheikh, Uday Sodhi, Aabhas Khanna and Maria Nigam. Full throttle ahead 🏎️ Link to the article: https://lnkd.in/gj-DXiSA

228

19 Comments

Thrilled to back Dima Shvets and Alexey Moiseenkov on the launch of Mirai On Device AI. If you’re building AI-native products, this one’s for you. Mirai is an on-device AI SDK for iOS and macOS that lets you run high-performance models directly in your app with zero latency, full data privacy, and no inference costs. So instead of sending tokens to the cloud (and paying for it), you keep things local. Even better, you don’t need a dedicated ML team or weeks/days of infra setup. One developer can integrate inference, routing, and optimization in minutes. Key features: — Supports Llama 3.x, Qwen 3, Gemma 3, DeepSeek-R1 — Bring-your-own-model support — Tools for fine-tuning, evals, structured outputs & fallback — Integration in a health tech product with 100M+ MAUs Built by the team behind Reface & Prisma. Backed by scout funds of Sequoia Capital & Index Ventures, and angels from OpenAI, Hugging Face, ElevenLabs, Unity and us, at NP-Hard Ventures 🤩 On-device AI will be a major unlock over the next few years, and help us to shift the way people interact with technology. And Mirai is early, ambitious, and razor-focused on enabling this shift. Go try it out! Link to SDK and GitHub documentations in comments. NP-Hard Ventures // Paul Veugen // Micha Hernandez van Leuffen

72

13 Comments

🔥 Excited to announce our ZenML + OpenPipe Integration for LLM Fine-Tuning in Production! After meeting the incredible OpenPipe team at a conference in New York last month, I was inspired by their vision for making LLM fine-tuning accessible. What began as an exciting conversation over coffee has now transformed into an integration that we're publishing live today. The challenges of fine-tuning LLMs for production are real - from data preparation complexity to reproducibility issues. Our ZenML + OpenPipe integration addresses these pain points by combining: • ZenML's production-grade MLOps orchestration • OpenPipe's specialized LLM fine-tuning capabilities ✅ End-to-end reproducible pipelines for fine-tuning ✅ Automatic deployment of models as soon as training is completed ✅ Complete lineage tracking from data to deployment ✅ Simple retraining as your data evolves We've detailed the entire implementation in our latest blog post, including a worked example of using this to fine-tune a customer support assistant that outperforms generic foundation models in both accuracy and cost. For teams looking to move beyond prompt engineering to truly domain-specialized AI, I believe this integration opens new possibilities. I'd love to hear your thoughts or answer any questions! #LLMOps #MachineLearning #AIinProduction #FinetuningLLMs #MLOps

58

7 Comments

the last few weeks feels like a mini breakthrough for AI apps hiding in plain sight - partial autonomy, high agency models, and context eng vs prompt eng. are all incredibly important concepts a lot of these principles were apparent but were just recently articulated (mostly by Karpathy - legend!!). one is the concept of partial autonomy apps - autonomy sliders, app specific UI, and “keeping the AI on a leash” two is the idea of small, high agency models that know how to solve problems vs knowing answers via tool-use / reasoning / multimodality (“maximally sacrifices encyclopedic knowledge for capability”) three is context engineering vs prompt engineering - putting the right stuff into working memory, building around the “jagged intelligence” of models etc these should probably be the starting point for founders thinking through AI native apps Refs: Partial autonomy apps: [Karpathy @YC startup school] High agency mini models: https://lnkd.in/gthCXtMe Context Eng: https://lnkd.in/gha2i7BX

101

10 Comments

Today, we announced the open sourcing of our AI tools for Online safety at Roblox. We invest significantly in this area in building the tools and running it on our platform spanning Core, Edge, and Cloud to enhance safety for our player community. The need for improving Digital safety in the world of AI is substantial for everyone, and this initiative marks a significant step in that direction that everyone can benefit from. #AISafety #Roblox #DigitalSafety https://lnkd.in/g_bftZCb

59

3 Comments

It’s an exciting day today! After four years of working in relative stealth-ish, we’re excited to publicly announce the launch of Camber (formerly Juniper), and the $50M in funding we’ve raised from Andreessen Horowitz, Craft Ventures, Y Combinator, Manresa Ventures, and many others who are committed to improving the accessibility of care. I joined Camber for the same reason I stay at Camber: the tangible impact we deliver every day. To date, we’ve helped turn three clinics around from bankruptcy. We’ve helped clinics collect thousands and thousands dollars more — we’ve seen them hire, open new clinic locations, and provide more care in our space as a result. We now power billing activities for ~90,000 patients across 42 states, getting our providers paid with no denials 93% of the time, and helping them collect 98% of the cash they’re owed over time. Honestly, that’s pretty awesome. The past four years have shaped so much of how I think about engineering. Engineering is cool, and building things is cool, insofar as it drives real impact in the world. I haven’t been more exhausted but more fulfilled than in the past four years. — Christophe Rimann, Celina Qi, Nathan Lee: Thanks for showing me this problem, and for burgers all those years ago. I guess the true partnership and back-breaking support is nice as well. Jade Shenker, Zack Zhu, Revan Sopher, Laura Skillington, Chris Maddox, Zachary Strenfel, Cayla Wanderman-Milne, Melanie Rogers, Chloe Oppenheim, Sy Yeu Joe C., Justin Sian, Mark Secada, Amy Zhang, Peggy Li: Thank you guys for taking a bet on us. I can’t imagine doing this without each of you. — We’re just starting! If making healthcare work sounds cool, reach out. I’d love to have you on our journey. https://lnkd.in/g8SjUq-j

212

16 Comments

GPUs have gotten faster, and models have gotten larger. But storage is still stuck in the past. Modern AI workloads need instant access to huge datasets. Instead, teams waste time on slow retrieval, idle GPUs, and costly pipelines. Archil is changing this. It delivers object-storage scalability at block-storage speeds. The platform allows for seamless, high-performance access to massive datasets. By eliminating cold starts and accelerating throughput 30X, Archil enables teams to fully utilize their infrastructure and move faster on training, analytics, and AI deployment. Archil’s founder Hunter Leath brings rare technical depth and operating experience from building Amazon EFS and optimizing Netflix’s cloud performance. This is a team building the foundational layer that AI workloads desperately need. I’m thrilled to lead Archil’s seed at Felicis with Nancy Wang and support their mission to reinvent cloud storage for the AI era. https://lnkd.in/gWueXhzz

65

2 Comments

Someone photographed me mid-selfie at TechCrunch Disrupt. Next to me, the 20 best startups of the event, incredible ones ⤵ Bleeding stops in 60 seconds. GPS-free navigation to 10cm. AI as an immune system for AI. One hits $1B in 3 years. The full lineup: → MacroCycle Technologies Upcycling plastic waste into virgin-grade resin at zero carbon → MIRAQULES Nano-biomaterials that stop bleeding within one minute → Nephrogen AI-powered platform discovering gene-delivery vectors for kidney diseases → COI Energy AI platform eliminating commercial energy waste → RADiCAIT AI generating synthetic PET images from CT scans → Mbodi AI  Teaching robots new skills via natural language in minutes → Skyline Nav AI | Pathfinder GPS-free navigation accurate to 10-100cm using computer vision  → Strong by Form Ultra-light timber composites replacing steel and concrete  → Pytho AI AI-powered mission planning tools for defense and military  → Glīd Autonomous, electrified road-to-rail freight transport  → Identifee Customer engagement platform for financial institutions with AI-powered CRM → CyDeploy AI-powered cybersecurity testing with digital twin sandbox environments → Charter Spacecraft engineering management software with automated insurance underwriting → Cyphr AI-powered small business lending platform analyzing cash flow for creditworthiness → Mappa AI voice analysis platform decoding human behavior for hiring decisions → Super Teacher Inc. AI-powered tutoring for elementary schools at $15/month  → CampusAI Polska AI education platform with virtual metaverse campus for upskilling workers  → UNLISTED Platform for buying homes before they hit the market (98% of inventory) → Unthread AI-powered support platform built into Slack for automated ticket management → Elloe AI | Immune System for AI AI safety platform serving as an "immune system" for bias, hallucinations, and compliance Which one are you betting on?

59

24 Comments

Qlay co-hosted Jeff Seibert, CEO of Digits (prev. sold Crashlytics to Twitter for ~$100M) and Jean-Denis Greze, ex-CTO of Plaid and co-founder of Town, which has raised $18M. Here's what I learned: 1. There’s no one-size-fits-all GTM strategy, especially in AI Should you license your AI model? Sell to players who already have customer bases? Or use the AI yourself and act as a service provider? This all depends on how the industry is structured and which business model opens doors to the most paying customers. 2. Brand is key, especially if you don't have huge traction Jeff asked Jean-Denis about the domain Town.com, which opened up a key insight: in a crowded market, your brand matters early. A simple, real-word domain can make your company look polished and credible, even before you’ve built significant traction. 3. Leadership is tested when things are hard Jean-Denis shared how a major acquisition deal falling through tested him as a leader. Getting the team re-energized and focused again after the setback was not an easy task. Thanks to everyone who joined us at DG717 Digital Garage for a high-signal conversation. DG717 Digital Garage · Tokumasa Yamashita · Digital Garage US, Inc. · Tom Nakata · Open Network Lab · Ethan Adam #Plaid #Digits #Qlay #Dropbox #Town #JeanDenisGreze #fintech #startups #techleaders #founders #entrepreneurship #angelinvesting #venturecapital #scalingstartups #ctoinsights #productleadership #seedfunding #earlystagefounders #startupjourney #startupgrowth #betweentwofounders #startupcommunity #DG717 #DGUS

52

▶️ We started Sandgarden after a year of brainstorming—with Britt Crawford, now our CTO, stubbornly saying “WPEngine for LLMs” on repeat. 🎵 Turns out that broken record wasn’t broken at all. It was a beautifully pressed 180-gram vinyl. And when we finally really heard what he'd been saying all along, it was music to our ears. Sandgarden was born. Like WPEngine simplified and scaled WordPress hosting for developers and businesses, we knew AI workflows needed the same treatment. The problem was clear: - PMs couldn’t spec - engineers got stuck doing grunt work, and - nobody trusted deployments. 🏋️ AI shouldn’t bottleneck teams—it should empower them to move faster, collaborate better, and deploy with confidence. That's what we're building at Sandgarden. 📩 Read the story of how it all began: https://lnkd.in/ghsuW3ws

50

4 Comments

Developers don’t need another “conference.” We need spaces where ideas spark, experiments happen, and community feels real. That’s why we’re bringing Camp AI by Auth0 to the Bay Area — think summer camp energy, but with the brightest minds in GenAI. You’ll hear from builders at Okta, CircleCI, Traversaal.ai, CodeAnt AI, and Acuvity, plus a panel with leaders from GitHub, Vercel, and Amazon Web Services.From late-night code chats to big-picture debates on how AI is reshaping the dev workflow — this is about building together, not just talking about it. And yes… there will be merit badges. Spots are limited → https://bit.ly/4lbaBol #CampAI #GenAI #DeveloperCommunity #BuildWithAuth0

75

2 Comments