Arjun Iyer

Yesterday was a big day at Fireworks AI where we announced our Series D: $1.5B raised at a $17.5B valuation. We've crossed $1B in ARR and now serve more than 40 trillion tokens per day. It still feels so early on this mission to build specialized intelligence—exciting times ahead! You can learn more here: https://lnkd.in/gri8Aytz

When we win competitive deals at Weave, one of the most common reasons we hear for why they chose us: our support is excellent. Here's how we approach it: I did every single support ticket myself until February this year, and onboarded every new customer too. Both were my responsibility because leading every ticket and running every onboarding call meant I knew exactly what confused people, what made them angry, and what they cared about. But once hiring and interviewing picked up and we started signing dozens more customers, I couldn’t do both. So we established a support rotation: every engineer takes a week as first point of contact for tickets now, and gets the same close look at customers I used to get alone. The rotation works, but every week an engineer spends on support is a week away from deep, uninterrupted engineering work. Scaling that trade-off further wasn’t sustainable, so we hired Wesley Lin as our first dedicated support engineer! During college, Wesley worked as a forward deployed engineer at a startup selling tech to restaurants. He had to solve technical problems and then explain the solutions to restaurant owners with zero technical background. Technical skill paired with the ability to make it make sense to anyone is exactly what we needed. Wesley is first point of contact now. The rotation still backs him up for anything that needs deeper engineering context. And support continues to be a competitive advantage!

I'm happy to share that I started a new position as Sr. Scientist, Study Manager at Merck! I am incredibly excited to work with the team to advance clinical development and improve patient outcomes. Special thanks to all my mentors, colleagues, and friends during my time at Regeneron who have helped me along the way. You will all be missed!

Everyone and their dog sells AI Firewalls. And they’re not even firewalls! How do you compare a SaaS one with an OOB API-based self-hosted container? I’ve called out AI Firewalls for applying archaic architectures to LLM use cases. User > Firewall > LLM > Firewall > User Very 2000 and late But there's more depth to these products. First, they're not firewalls because they don't filter packets. They host detection logic, so you have to re-learn how they fit in your env I found these 4 dimensions to evaluate them against: ➡️ Stateful - The firewall maintains its own persistent, immutable record of the entire interaction context (prompts, responses, tool calls, API calls, reasoning traces) across turns and sessions, independent of and unaffected by any truncation or summarization applied to the LLM’s context window. ➡️Out-of-band - firewalls are obviously inline, but you also have an out-of-band little brother that is used for monitoring mode. The analysis verdict can then be implemented in your application. ➡️Vendor-hosted - detection logic runs on the vendor’s infrastructure. You can access that via: 1) API endpoints exposed by the vendor 2) An SDK, where a vendor library or framework hook embedded directly in your application or agent code 3) Through the CDN/SASE, where the vendor implements the firewall in the rest of their as-a-service network. 4) Proxy, where a non-CDN vendor hosts the firewall and you route your prompts through their firewall. ➡️Self-hosted - vendor gives you a form factor (container or VM) that runs the detection and enforcement engine to deploy in your prefered environment. Let's see how these apply across vendors: Statefulness I wrote a paper for AccuKnox about their stateful firewall, which is a very imporant capability for evaluating intent. I also found Aiceberg can do full session-level analysis, but requires some configuration. WitnessAI says the right words, but not enough to confirm. As I was writing that paper, I saw that statefulness isn't always mandatory. LLMs are stateless so you always send the whole context. It matters once context gets summarized or truncated, because the firewall otherwise loses data the LLM no longer sees, which can contain malicious intent indicators. Vendor hosted Folks like Cloudflare, Cato Networks, F5, and Akamai Technologies run their own SASE/CDN, so they just put the firewall in their product. For an existing customer, this is a toggle switch and another line item on the invoice. Self-hosted HiddenLayer has some great docs that explain how this is architected. You can also look into A10 Networks, Inc and Check Point Software's Lakera, now part of Check Point. Palo Alto Networks is the most typical 'filter packets' firewall you'll see here, though detection runs in their cloud Out-of-band Nearly everybody does OOB verdict-only, because they can only enforce when in-line. There are OOB enforcement techniques like revoking tokens, but that's for another post

Thank you to everyone who joined us at the Friends of Clearwater Summer Party, an unforgettable evening on the rooftop at Sea Containers London in London. The night offered stunning panoramic views across the London skyline. And with Clearwater’s recent announcement of its $8.4 billion take-private acquisition by Permira and Warburg Pincus LLC, there was even more reason to celebrate with our friends. Guests enjoyed hand drawn portraits from a live illustrator and captured keepsake snapshots from our photo booth. Our nitro ice cream stand was a highlight, offering partygoers not only a cold treat on a hot summer’s night, but a theatrical experience, with fog drifting over the counter as the chef freshly prepared each order. We can’t wait to welcome our friends back to our next flagship EMEA event, Clearwater Connect London, on 21 October.

On the floor at the Fixed Income Leaders’ Summit in Boston, one challenge emerged from Clearwater’s roundtable discussions on workflow and data. Leaders were not short of AI ambition but found a data gap was holding them back. Industry experts pointed to data fragmentation as the biggest barrier to AI adoption in fixed income and agreed that clean, reconciled data is needed to make accurate decisions. Read what fixed income traders, portfolio managers and trading technologists discussed: https://lnkd.in/eDeEEZ9Z

I’m excited to announce that I’ve joined Lorimer Capital as Director of Credit & Underwriting, where I’ll be leading the evaluation and execution of construction and bridge loan investments across Lorimer’s real estate credit platform. A big thank you to my colleagues, mentors, and friends at Greystone and Columbia Business School, whose guidance and support shaped the path that brought me here. You’ve helped me broaden my perspective and deepen my thinking over the last several years, and for that I’m deeply grateful. I look forward to helping build on Lorimer Capital’s momentum as the firm continues to expand its real estate credit business. https://lnkd.in/g5HvHtcs

Insurers leave an estimated $15 billion in subrogation on the table every year (NAIC). Part of why: in the data we see, 80% of documents entering claims come in classified as "miscellaneous." This week at the Future of Insurance in Chicago, the Scale AI team demoed how agentic AI changes that math: connecting distribution, claims, and underwriting into one connected enterprise. Surface the recoverable dollars. Replicate your best adjusters' rigor across the book. Flag fraud patterns earlier. Improve the loss ratio. Subrogation is where we start. The connected enterprise is where it goes. #Insurance #AgenticAI #InsurTech #Claims Thanks to Yonah Broth , Mostafa Shehadeh and Tejas Pathak for amazing teamwork.

I am honored to have been invited to join the Board of Directors of SpaceX. I look forward to the opportunity to work with Elon again. In addition to SpaceX, I continue to serve on the Boards of leading public companies: Block, Ethos, MongoDB, Natera, and Unity Software. I also work closely with the visionary founders of several emerging private companies. Given the amount of time that I need to devote to these companies, and my own continued new investment activity, I recently concluded my transition from Sequoia Capital. Over my nearly 24 years at Sequoia, I was fortunate to lead investments in iconic founder-led companies like YouTube, Instagram, Block, MongoDB and Natera. Beyond just leading investments, what brought me joy was building long-term relationships with founders and management teams, and helping with the arc of company building. Similarly, what I’m most proud of about my time at Sequoia was the way in which we developed our distinct culture, thoughtfully recruiting and nurturing a next generation of investors. We leveraged each other as we pursued opportunities, and we made investments as a team, ensuring that all of us felt ownership and accountability for our decisions. Sequoia is a truly special firm, and I know that with its winning team, the partnership’s future is bright. I am excited for what lies ahead for me next, driven by an enduring passion to support  founders as they build category defining companies.