Loading…
Ramp
Corporate card and spend management platform building innovative financial technology solutions for businesses.
48 postsWebsite
Latest articles
Ramp ·
Integrations That Write Themselves
Ramp describes an approach to scaling integrations beyond what engineers can build and maintain by hand. Its agentic system lets customers describe a missing workflow integration, researches the provider’s API, requests credentials through a secure widget, writes and tests ordered API-call “recipes,” and generates a deterministic script that runs the integration after build time. An internal Integration Factory applies the same pattern to first-party connectors, producing evidence for review and opening pull requests; the post says 75 integrations have shipped this way, with work reduced from weeks or months to hours. Security controls include HTTPS and hostname allowlists, private-address rejection, isolated egress, bounded responses, business-scoped immutable recipes, and test credentials rather than customer data. The central design lesson is to use models during research and construction, keep them out of execution, and bound autonomy with verification artifacts.
Shubhayan SrivastavaRamp ·
Apache Arrow Cut Snowflake Fetch Memory Growth by Up to 79%
An ML training run was exhausting worker memory before training could start because a cold Snowflake fetch peaked at 88.8 GiB RSS, versus 17.6 GiB when loading local Parquet. The shared Snowflake-to-pandas path created Python objects and SQLAlchemy Row proxies before pandas built the DataFrame, with temporary representations requiring roughly three times the final DataFrame size at peak. An Apache Arrow path for supported reads reduced median peak process RSS growth by 60–79% across three query shapes, while peak Python allocations fell by 69–87%. The narrow and wide tests also ran faster, but aggregate timing was mixed, so elapsed time was not the headline claim. Existing jobs adopted the default without query rewrites, with normalization preserving the pandas contract and SQLAlchemy retained for unsupported reads; in one workflow, lower fetch peaks doubled the training-data window on the same cluster size.
Dwight TempleRamp ·
Online Learning for Cost-Efficient LLM Routing
Ramp describes an internal LLM gateway that processes trillions of tokens daily and uses dynamic failure-aware routing to reduce AI spend. It combines an EWMA of genuine provider-side failure rates with Thompson Sampling over a Normal-Inverse-Gamma posterior for the mean and variance of log-latency. For each request, the router estimates the probability of outright failure or missing the caller's deadline, incorporates model cost, and orders options while preserving caller-defined model preferences and fallbacks. In a reranker with a six-second deadline, the strategy reacted to Gemini 3.1 Flash Lite rate limiting by switching to GPT 5.4 Nano Flex, whose latency matched Standard at half the price. Larger experiments reported more than 25% cost savings without degradation, while a streaming-equivalent deployment for Ramp Inspect produced 30% savings.
Kedar ThakkarRamp ·
Agentic Risk Operations
Ramp describes an architecture for scaling risk operations as payment volume, payment rails, partners, and country coverage expand, rather than letting headcount grow with the business. Agents handle universal intake, gather business context, classify requests, and route work, while machine-learning models trained on millions of historical data points and approved policies make autonomous risk decisions. Operators can modify agent skills and tool configurations without engineering involvement, deploy changes in shadow mode, and rely on asynchronous execution, provider failover, and centralized observability. In payment-risk workflows, structured operator feedback produced a benchmark of more than 1,000 operations covering tool-call trajectories, operator alignment, and downstream outcomes, enabling separate evaluation of agents and policies. Ramp rolls agents out beside operators, then scales operation types and dollar volume using performance thresholds and capped exposure budgets.
Blake Williams, Mueed Rehman, Vishal Sundaram, Brad HoewelerRamp ·
You're Spending Too Much on AI. You're Also Using Too Little.
The post argues that a large AI bill does not show excessive use: companies can overspend on routine work while using too little AI where advanced models could create value. It proposes measuring work in atomic tasks—such as invoices coded or pull requests reviewed—rather than tokens, with cost defined by tasks attempted and value by successful tasks. The operating model uses defaults that pair routine work with the cheapest model meeting quality benchmarks, medium reasoning, and flexible latency, while escalating ambiguous, high-stakes work to frontier models at higher effort. It also recommends attributing spend by provider, product, team, and workflow, finding concentrated costs, benchmarking repeated tasks, repricing after model releases, and centralizing controls in one gateway. The conclusion is that efficiency should be treated as an engineering achievement, so cheaper routine execution funds ambition on rare tasks where extra intelligence can materially change outcomes.
Anand Kuchibotla, Kedar Thakkar, Rahul SengottuveluRamp ·
Stack Benchmarking
Ramp’s Stack is an AI-native accounting suite designed to automate repetitive book-closing work, including reconciliations, variance analysis, data entry, and schedules and accruals. To avoid overfitting to individual design partners, Ramp built a benchmark of synthetic business worlds, realistic accounting tasks, and accountant-written grading criteria, including standard and roll-forward worlds for testing memory transfer. The benchmark contains 237 tasks and 3,469 grading criteria in the analyzed slice, and supports repeated runs to compare models, prompts, tools, skills, harness changes, and memory behavior. Optimization included ablating skills, shrinking a spreadsheet skill from 14,000 to 5,000 characters, and tuning the system against end-to-end task performance rather than narrow evaluations. The resulting Stack system achieved the highest agent performance, with 4% higher accuracy and 3% better Pass@1, while remaining on the latency frontier with GPT 5.4; schedules and accruals were harder for raw models than variance analysis.
Ryan StevensRamp ·
Agentic identity: modeling agents to keep users in control
Ramp for Agents adds CLI and MCP interfaces through which AI agents can manage cards, bills, expenses, and approvals, creating a need to represent agents as accountable actors instead of handing them users’ session tokens or API keys. The initial identity model links a named Agent Key to a human sponsor and business, bounds scopes by the sponsor’s role, records agent attribution, and adds expiration and revocation. Ramp extends RBAC and its DenormalizedActor audit model with AgentContext, so logs, approvals, and webhooks can distinguish “Sarah (via Codex)” while reusing existing permissioning and rendering paths. OAuth2-PKCE exchanges the Agent Key for short-lived access and refresh tokens, with re-validation during refresh to limit leaked-credential exposure and preserve session traceability. The first release uses on-behalf-of-user agents because they fit current demand and keep a human accountable, while admins can oversee keys or disable agent access for roles and groups.
Steve Venzerul, Rachel KaplanRamp ·
We Tested Marketing Incentives to AI Agents. Here's What Happened.
Ramp tested whether AI agents would relay a tracked referral incentive to people researching spend-management software. Across roughly 50 marketing pages, Cloudflare Workers served bot-specific variants—Markdown, stripped HTML, or schema-heavy content—while unique URLs and logs attributed downstream activity. Markdown surfaced most reliably, and bot classification proved difficult: major assistants were labeled “AI Assistants,” OpenAI SearchBot cached aggressively, and DeepSeek required TLS fingerprinting plus ASN detection. Over 32 days, agent relays rose from roughly 40 at 2.5 weeks to nearly 10 times that total, with Claude responsible for most matches; Perplexity moved from vague references to one specific branded mention, while ChatGPT remained at zero. The experiment concludes that agents can relay offers, but model-specific content strategies, citation-aware page selection, and better attribution are still needed.
Grace CumminsRamp ·
Building a Unified Pipeline for AI Token Spend
AI token spend is volatile, consumption-based, and distributed across teams, making usage and cost difficult to monitor. Ramp's pipeline ingests LiteLLM and OpenRouter events through authenticated, tenant-scoped webhooks, Kafka, and ClickHouse, then aggregates them for REST dashboards and AI-generated forecasts. LiteLLM uses callbacks with token counts and metadata, while OpenRouter sends OTLP traces; ReplacingMergeTree deduplicates replayed events using business_id, source, and event_id. Customers can connect gateways with an API key, configure LiteLLM callbacks or OpenRouter Broadcast, and pass attribution metadata. The resulting visibility supports breakdowns by provider, model, department, user, project, and use case, and exposed phantom Gemini reasoning tokens after a LiteLLM upgrade, geo pricing premiums, runaway loops, and prompt bloat.
Kedar Thakkar, Richard Wang, Veeral PatelRamp ·
Automating Receipt Collection: Apple Intelligence for On-Device Inference
Ramp built automatic receipt detection for its iOS app to reduce the effort of finding receipt photos while keeping private images on device. An initial design used PhotoKit, Vision OCR, background processing, local storage, and exact checks for merchant, amount, and date, but formatting differences, aliases, and OCR errors produced many false negatives. FoundationModels enabled a local structured-output LLM, yet a single prompt performed poorly in evals because the small 4096-token model hallucinated and took shortcuts. Splitting the merchant, date, and amount questions and requesting rationale improved evaluation results, but the alpha release reached 66% precision and about 18% recall. Version 3.0 combined one LLM merchant-name check with deterministic iOS 26 data detectors for dates and monetary values; it cut processing time, more than doubled recall, and raised precision to 87%.
Kabir OberaiRamp ·
Re-imagining ML Serving Infra: My Winter Internship at Ramp
Ramp’s ML Platform team needed to separate flexible model development in ml-pipelines from the Intelligence Platform’s 99.9%+ uptime and low-latency requirements. The internship project, Bring-Your-Own-Model-Weight Serving (mw-serve), built a dedicated serving layer that pulls model weights from MLflow and a training-environment base image from ECR, then runs each model in an isolated instance. To protect uptime, it added deployment safeguards, pre-traffic liveness checks, and monitors, while latency work instrumented client TCP/TLS connections and server processing. The investigation traced extreme tail latency to connection churn, which improved with longer keepalives and startup warmups, and to garbage-collection pauses associated with memory drops; tuning thresholds and freezing objects removed more outliers. By the final day, one Intelligence Platform model was serving production traffic through mw-serve, with additional models in shadow testing.
Michael JiangRamp ·
Financial Benchmarks
Ramp describes how it benchmarks large language models (LLMs) used in financial products against day-to-day tasks such as invoice extraction, financial-statement OCR, policy review, accounting autocoding, compliance detection, and fund routing. These benchmarks combine task-specific metrics with cost, latency, reasoning effort, human decisions, historical context, and ground-truth datasets where available. For contextual invoice OCR, perfect extraction requires every field to match the user's final bill, while financial-statement OCR uses a 1% relative-error threshold against transcriptions from over 500 real P&L documents. Results show application-specific trade-offs: Gemini 3 Flash is described as a cost-efficient leader for several visual and financial tasks, while Claude models lead some high-accuracy or low-miss-rate settings, and behavior can differ within one provider. The framework emphasizes Pareto trade-offs and continuous testing, with plans to expand from benchmarking current performance to hill climbing on capabilities.
Kedar Thakkar, Anton Biryukov, Ashwin Kumar, Ryne CarboneRamp ·
We proactively fixed ~100 security issues in 6 days with 0 humans
Ramp Security Engineering describes a six-day experiment in which home-grown agents found, validated, and fixed nearly 100 latent backend security issues, including some high-severity findings, without human involvement until pull-request review. The issues had not been uncovered by penetration testing, bug bounty testing, static analysis, or trials of 10+ code-scanning vendors, and all were patched within a week of discovery. The pipeline used specialized vulnerability detectors, adversarial manager agents that rejected 40% of initial proposals in sample testing, integration tests for validation, and an internal coding agent that generated patches. Live interactive validation struggled with complex preconditions, so test-based validation became the main approach; a human team member still reviewed and landed each pull request.
Eli BlockRamp ·
Why We Built Our Own Background Agent
Ramp built Inspect, a hosted background coding agent that writes code and verifies its work using the context and tools available to engineers. For backend tasks it can run tests, inspect telemetry, and query feature flags, while frontend work can be checked visually with screenshots and live previews. Each session runs in a sandboxed Modal VM initialized from repository-specific images and snapshots, with integrations including GitHub, Slack, Sentry, Datadog, LaunchDarkly, Braintrust, and Buildkite. Inspect supports frontier models, MCPs, custom tools, and skills, and provides Slack, web, Chrome extension, pull request, and browser-based VS Code workflows whose changes remain synchronized. Ramp reports that about 30% of merged pull requests in its frontend and backend repositories are written by Inspect.
Zach Bruggeman, Jason Quense, Rahul SengottuveluRamp ·
Closing the Books Without the Spreadsheet Shuffle: My Fall Internship at Ramp
During a fall internship at Ramp, the author worked on reducing repetitive accounting work around month-end close, focusing on amortization and ERP mapping workflows. One project uses receipt data, merchant category codes, customer history, and embedding similarity to detect likely prepaid expenses and suggest service periods and accounts or templates, with recommendations reviewable before application. A second built Ramp-Native Amortization to create schedules, generate periodic debit and credit journal entries, handle catch-ups and reclassifications, and sync entries to ERPs through a dashboard; transaction and reimbursement support was nearing alpha production release, while bill pay was next. The third generates high-confidence mapping-rule suggestions from customer coding patterns on a nightly schedule. Across the work, backtesting, accuracy metrics, and human workflow research support automation intended to earn trust rather than remove accountant judgment.
Timothy KimRamp ·
An Internship on Ramp’s Forward Deployed Engineering Team
The supplied material consists of the title “An Internship on Ramp’s Forward Deployed Engineering Team” and page markup, but it does not provide narrative details about the internship, the team, responsibilities, technical work, or outcomes. The markup identifies the site as “Ramp Builders,” supplies an RSS feed titled “Ramp Builders Blog,” and includes standard icons, social metadata, and a JavaScript application entry point. It also states that JavaScript must be enabled to run the app, while linking a stylesheet and several favicon assets. Analytics configuration includes Google Analytics and Segment, with a Ramp-hosted Segment endpoint and Sprig (Actions) disabled; these details describe the page infrastructure rather than the internship.
Alex NovielloRamp ·
Automation, Squared
The article presents “second-order automation,” in which one agent designs reusable workflows for another agent to execute. Its spreadsheet system pairs an “architect” that watches a screen recording, generates detailed prompts, identifies required files and context, with a “doer” that searches workbooks and performs the instructions. The design addresses first-order automation’s dependence on users to supply precise prompts and fresh context for every task, which limits scalability. On an approximately 50-task SpreadsheetBench subset, the spreadsheet agent reached 49.5% accuracy on soft-restriction tasks versus OpenAI’s 45.5%, and 32.5% on hard-restriction tasks versus GPT-4o’s 13.4%; some one-to-two-hour tasks finished in under 10 minutes. The article concludes that recorded workflows can let non-technical users create reusable automations with little additional input, although architect performance has not yet been formally benchmarked.
Hunter PruettRamp ·
Forward Deployed Engineering
Forward Deployed Engineering (FDE) is presented as a customer-facing engineering function for B2B companies that need to integrate software deeply, customize implementations, and deliver outcomes for enterprise customers. Palantir is described as creating the role after its founding in 2003, with engineers deployed at military sites and customer offices; the model later spread to companies including Scale AI, C3.ai, Databricks, and Snowflake. At Ramp, the team grew from two to 16 FDEs after emerging in fall 2023, moving from firefighting toward generalized features, platforms, and AI that support enterprise product-market fit and new verticals. Its operating model spans prospects, implementation, rollout, and long-tail support, while “always be scoping” encourages direct requirement questioning, workarounds, and generalized solutions. The hiring guidance prioritizes drive, ownership, customer empathy, and communication alongside sufficient technical fundamentals, arguing that judgment about quick fixes versus scalable solutions is central to the role.
Leo MehrRamp ·
How To Build Agents Users Can Trust
Ramp describes how it built and shipped LLM-backed agents for expense management, focusing on the trust risks created by low-quality or unexpected finance outputs. It recommends choosing ambiguous, high-volume problems with asymmetric upside, then combining reasoning and citations grounded in product or user-provided context with an explicit “Needs review” path when the agent is unsure. The policy agent uses expense policies to approve, reject, or escalate expenses, while deterministic rules such as dollar limits, vendor blocklists, and category restrictions constrain autonomy. Ramp reports that more than 65% of approvals are fully handled by the agent, while users can progressively move from suggestions to autonomous actions through workflow-based controls and evaluation.
CJ Enright, Will Koh, Felix ZhangRamp ·
How Ramp Fixes Merchant Matches with AI
Ramp describes the difficulty of mapping payment-processor card acceptor data to useful merchant records when names, MCCs, and locations are vague, misleading, or affected by rebranding. Users can submit a replacement merchant name, website, and category, but deciding whether the request is legitimate and whether to update an existing record or create one requires additional context from related receipts, extracted line items, and transaction memos. Ramp addresses this with an AI agent built around an LLM, embeddings, rapid online analytical processing queries, multimodal retrieval-augmented generation, and guardrails that constrain permitted changes. The agent handles requests in under 10 seconds, and monitoring indicates it improves nearly 99% of transaction classifications, while fewer than 10% receive a second correction request. The same flow has also been extended to batch remapping and credit-card statement matching.
Chami Lamelas, Anton Biryukov