Teddy Citrin

Today, we're releasing Ramp CLI to let agents manage your company's finances. 50+ tools across cards, bills, expenses, travel, and approvals. Fewer tokens than MCP, and comes with pre-built skills like receipt compliance and agentic purchasing.

Over 1 billion sandboxes have been launched on Modal. Since launching three years ago, we've seen Modal Sandboxes become foundational to how AI is being built. Today, teams like @Lovable, @tryramp, @cognition and more are using Modal Sandboxes to power everything from coding platforms and background agents to RL infrastructure at scale.

Today, we're launching Ramp Agent Cards. There's been no safe way for agents to spend money, until now. Ramp Agent Cards give agents the ability to spend, governed with real spend limits, merchant controls, and full visibility into every transaction.

The FA4 paper is finally out after a year of work. On Blackwell GPUs, attention now goes about as fast as matmul even though the bottlenecks are so different! Tensor cores are now crazy fast that attn fwd is bottlenecked by exponential, and attn bwd is bottlenecked by shared memory bandwidth.  Some fun stuff in the redesigned algorithm to overcome these bottlenecks: exponential emulation with polynomials, new online softmax to avoid 90% of softmax rescaling, 2CTA MMA instructions that allow two thread blocks to share operands to reduce smem traffic.

We partnered with @mercor_ai to post-train custom models on high-quality expert data from fields like law, investment banking, and consulting. Our latest model ranks #1 on the APEX-Agents leaderboard in corporate law and #4 overall. Domain-specific post-training on high-quality, organization-specific data can systematically close the gap between general AI competence and expert-level reliability, making capable enterprise agents practical and affordable for knowledge-intensive industries. appliedcompute.com/case-studies/m…

Introducing Devin 2.2 – the autonomous agent that can test with computer use, self-verify, and auto-fix its work. Try it for free! We’ve also overhauled Devin from the ground up: - 3x faster startup - fully redesigned interface - computer use + virtual desktop ...and hundreds more UX and functionality improvements.

Companies should own their data and intelligence, instead of renting it. That’s the thesis behind Applied Compute. So what does that mean in practice? When we help companies turn their latent knowledge into specialized proprietary agents, each customer deployment is fully isolated and runs inside that customer’s VPC. The point is for customers to own their data and the intelligence derived from it, so nothing needs to leave their environment. They can also self-serve whenever they want. Beyond that, diffusion concern also assumes implementation for one company maps cleanly to another. But what makes enterprises unique is messy context spread across legacy systems, teams, workflows, and years of accumulated data. Done properly, deployment ends up being tailored to those constraints, so you can’t copy it to a competitor and get the same result.

DoorDash has learned a lot about shipping impactful AI products through our partnership with @ypatil125 @rhythmrg @lindensli at the @appliedcompute team. We're already seeing additional traction collaborating them in other use cases we hope to share soon.

Excited to finally share this! It was an amazing collaboration with @andyfang and the @DoorDash team! We’re thrilled to continue partnering with one of the most innovative and execution-focused AI teams in the world.

Eric Glyman @eglyman runs @tryramp, which now powers >2% of US corporate spend. @arampell and I dug in with him on why financial systems have proven resistant to automation, the second-order effects of AI on business, and Ramp’s strategy.

There are two non-negotiables in accounting: the books must be correct, and they must be ready on time. For decades, companies have satisfied those constraints through an extraordinary amount of manual effort. Highly trained professionals code transactions, re-approve familiar expenses, reconcile mismatches after the fact, and compress all of it into the ritual of month-end close. It works. But it is fundamentally retrospective. Today, @tryramp is introducing an Accounting Agent designed around a different premise: what if bookkeeping happened as the business operated, rather than after it? The agent captures, codes, reviews, validates, accrues, and reconciles spend continuously. It learns directly from the people who understand the nuances best, the accounting team itself, and applies that context in real time. At @perplexity_ai, where velocity is part of the company’s identity, this has allowed their team to stop choosing between speed and accuracy. The majority of transactions are now coded automatically while remaining audit-ready, enabling close to start on day one instead of day thirty. What’s been most striking is how the system learns the subtle, company-specific logic that historically lived only in human judgment. As Jim Romano, CFO at @statesidevodka, described it, the agent is already identifying patterns like when spend belongs in samples rather than travel and entertainment — the kinds of decisions that typically require institutional memory. As he put it, the goal is simple: finance teams should focus on exceptions, not the easy stuff. We’re also seeing the second-order effects emerge quickly. Teams report spending dramatically less time reviewing transactions and substantially more time on planning, analysis, and growth. As one CFO told us, “What used to take hours of manual review now happens automatically. I’m spending nearly all of my time thinking about where the business should go, not retracing where it’s already been.” There is a broader shift underway in accounting. The central question is moving from “what parts of close can be automated?” to “should close even be a discrete event at all?” One belief that increasingly guides our work at Ramp is that information latency inside companies is an invisible tax. When financial truth lags behind operational reality, organizations make slower and often worse decisions. As transaction data becomes inherently digital and systems become capable of learning institutional context, continuous close stops being aspirational and starts becoming inevitable. One thing that surprised us while building this: accounting isn’t constrained by a lack of rules — it’s constrained by how many of those rules are unwritten. Much of financial operations lives in patterns that experienced teams simply know. Seeing software begin to absorb and apply that tacit knowledge has been one of the clearest signals that accounting is entering a new phase. Accounting has always been the record for business reality. Our goal is to help it become something closer to real-time truth. Proud of the team, and grateful to the customers building this alongside us.

The real bank for real Americans doing real things! The cutting-edge business model is based on next-gen concepts like "don't let your client's money disappear", "care at all about national security", and most importantly, "the market sometimes goes down".

Decagon has raised an additional $250 million, tripling our valuation to $4.5 billion in less than six months. We’re on a mission to help every company create concierge customer experiences. We believe every customer deserves support that is personal, proactive, and always there. If you’re the customer of something, it should feel like you’re the only one, 24/7. This round was co-led by new investors @CoatueMGMT and @IndexVentures, with participation from @Chemistry, @Definition_Cap, and @StarwoodCapGrp. It also includes continued support from @a16z, @A_StarVC , @Accel, @AvraCap, @BainCapVC , @EladGil, T.Capital, @ForerunnerVC, and @RibbitCapital. Read more on the raise and what we’re building next in our blog below.

Meet Brian. Brian’s been carrying accounting on his back for a long time. Super Bowl Sunday, he finally gets backup.

@thejessezhang and @ashwinsreenivas are rebuilding how businesses talk to customers with @decagonAI.

@kath_mcmahon and @velvetatom are pushing biological intelligence to new places with @ValthosTech.

Most AI review tools today center around asking an arms-length agent to catch & report potential bugs. This is really valuable! But until we reach the point where you can confidently hit "Merge" on a 5000-line agent PR, you're still bottlenecked on reviewing the code yourself. This will stay true for a while even as the tools get better. Would you rather have an arms-length AI that catches 80% of bugs or an AI-powered review UX that makes *you* 5x faster? Probably the latter since you'd still have to review the whole PR yourself to catch the last 20%. Of course, the best review experience should have both! We built Devin Review with these thoughts in mind. Let us know what you think!