BreakpointRamp describes OCA, an AI on-call assistant that investigates incidents, recommends next steps, communicates through Slack, answers follow-up questions, and prepares code fixes.
Ramp engineering blog
The post explores integrations that can be generated and maintained autonomously. It highlights an agentic approach to reducing the manual effort required to build integrations.
The post explains how replacing row-by-row Snowflake result materialization with Apache Arrow reduced fetch-memory growth by up to 79%. The change doubled the usable training-data window for an ML workflow without increasing cluster size.
The post examines online learning for routing LLM requests across models. It focuses on using adaptive selection to improve routing efficiency and cost.
Ramp describes its plans for building agentic risk operations and the direction of its program.
The authors argue that high AI spending often reflects inefficient purchasing rather than excessive usage, while organizations may still be underusing AI. The post appears to examine the tension between AI costs and adoption value.
Ryan Stevens explains how Ramp built an accounting benchmark for its agentic accounting assistant, Stack. The post examines frontier-model performance, skill ablation, and memory design.
The post introduces agentic identity as a way to model software agents acting on users’ behalf. It explains how this approach can help finance teams adopt agents while preserving user control in increasingly autonomous organizations.
Grace Cummins describes an experiment testing tracked marketing incentives aimed at AI agents on Ramp webpages. The post summarizes what the team learned about reaching agent-driven audiences.
The post describes Ramp's effort to build a real-time pipeline for tracking AI token spend and usage for finance teams.
Ramp engineers describe using Apple Intelligence to match receipts in a user’s photo library with transactions. The approach performs inference entirely on-device, emphasizing privacy and local processing.
Michael Jiang recounts an internship project re-imagining machine-learning serving infrastructure at Ramp. The post highlights containerizing models for a critical production service.
Ramp compares large language models on real-world financial tasks and explains why production-oriented benchmarks are more useful than generic evaluations. The post highlights how task relevance and operational context can improve model assessment.
Eli Block describes a six-day experiment in which autonomous agents identified, validated, and patched roughly 100 security vulnerabilities without human intervention.
Ramp engineers describe why they built a custom background coding agent. The post examines how such an agent could accelerate software development.
Timothy Kim recounts a fall internship at Ramp focused on building accounting automations to streamline finance teams’ month-end close and reduce spreadsheet work.
The post recounts Alex Noviello’s internship on Ramp’s Forward Deployed Engineering team, including work on automatic hotel receipt retrieval. It offers a high-level view of the internship and engineering problem without technical implementation details.
Automation, Squared explores how software agents can learn from humans performing real work, using a single recording as the basis for scalable workflow automation.
The post examines forward deployed engineering as a technical role and discusses its place in enterprise-focused engineering organizations.
The post discusses lessons from building an agent to automate expense approvals. It focuses on designing agent-driven workflows that users can trust.