← All speakers

Alexander Shevchenko

Conference affiliation: Ramp

Alex Shevchenko heads applied research at Ramp and leads Ramp Labs. His work includes Ramp Tour Guide, an assistant that operates Ramp’s interface, and Ramp Sheets, a spreadsheet editor whose formulas and cells let finance professionals inspect an agent’s work. Across these projects, he has worked on both how agents act and how people can assess their results.

On this page

From process mining to interactive agents

Before Ramp, Alex worked as a software engineer in process mining at Microsoft. His independent software includes an inverted index written in Rust, which uses BM25 ranking for search, and Gondola, a moodboarding application built with Rust and WebAssembly.

Alex joined Ramp around 2023. He initially worked on AI infrastructure, including CUDA drivers and memory partitioning for model services, before moving into experimental products. He later established Ramp Labs with product lead Alex Stauffer. The group explores AI without a predetermined product problem or quarterly metric, while looking for ways its experiments can contribute to Ramp’s software.

Alex led Ramp Tour Guide, an assistant that navigates the finance platform through clicks, scrolling and text entry. Users can watch its actions, read explanations and take control. The team found that generating the next action from the current application state worked better than planning an entire tour in advance. It also combined planning and action generation into one model call after a two-call approach proved too slow. Alex’s prescription was concrete: simplify the page representation and give the model a clearly labeled set of available interactions.

Making spreadsheet automation inspectable

Ramp Sheets grew out of an effort to document accountants’ month-end work. The team explored process mining—reconstructing workflows from event logs—and processed recordings into documentation. It then tried generating Python and workflow automations directly, but finance staff found the code difficult to verify.

Much of the recorded work already happened in Excel. The team shifted toward generating spreadsheet modifications that accountants could inspect through familiar formulas and cells. For the public product, it added a text interface so users could request a model without first recording an established process. Alex describes this progression in an interview about building Ramp Sheets and Ramp Labs.

Ramp launched Sheets in November 2025. Its agent can follow formulas across worksheets, clean data and build financial models. An Excel evaluation service applies formulas, reads their results and sends updates to the editor for recalculation and display; a code interpreter handles calculations that cannot be performed in Excel itself. Ramp’s finance and product colleagues reviewed early outputs against their own spreadsheet practices, helping the team assess whether it was producing workbooks a finance professional could use. Alex and Ramp colleague Will Koh explain that development in Ramp’s account of the project.

Improving the work before frontier models catch up

Alex favors adopting stronger general models while researching limitations that still cost users time. In his 2026 talk on Ramp Labs, he argues that waiting for the next model leaves useful improvements unbuilt. Three projects illustrate where his team has chosen to intervene:

  • Give a frontier planner a specialist. When users supplied several spreadsheets and requested a discounted cash flow model, Opus 4.6 could spend considerable time investigating the source material. Ramp partnered with Prime Intellect to post-train a 35-billion-parameter Qwen 3.5 model for financial question answering. Alex reported that it surpassed Opus 4.6 on Ramp’s task-specific accuracy measure at latency comparable to Haiku. The frontier model continued to plan the analysis, handing narrower financial questions to the specialist. He described the system as deployed in production.
  • Prepare predictable context ahead of time. Alex’s exploration of KV cache cartridges starts with month-end close: accountants can anticipate needing the ledger, aging report and related records. Preparing a cartridge overnight could let an agent begin with that context already loaded, rather than reconstructing it through retrieval in the morning. Alex said the techniques could compress 20 to 30 times as much context into the cache. He also acknowledged that model providers might eventually offer the capability through their APIs; he values the experiments partly because they teach the team how to use it when that happens.
  • Check what parallel agents discover. Ramp Labs ran 10,000 security-testing agents in parallel, then used models to deduplicate their reports and adversarially attempt to reproduce the findings. Alex reported seven patched high-severity vulnerabilities. One was an email-triggered Lua injection that required tracing interactions among components that were individually safe. The work also exposed a cost tradeoff: cheaper open-weight models could have lower detection rates individually, yet remain useful when their lower token cost allowed more parallel exploration.

1 conference talk

Loading more talks and highlights…

References