Ramp’s Case for Specialist Models and Prepared Context

Alexander Shevchenko of Ramp explains why an application company still does applied AI research instead of waiting for stronger general models.

Specialize one bottleneck

Ramp Sheets keeps a frontier model as the planner but delegates financial question-answering to a smaller specialist.
Ramp and Prime Intellect post-trained a Qwen 3.5 model for that task.
Shevchenko reports accuracy above Opus 4.6 at Haiku-level latency on Ramp’s financial question-answering task.
The talk gives no public evaluation protocol or exact latency figures; this is a company-reported component result.

KV cache cartridges package context into a model’s stored working state ahead of time.
Shevchenko describes packing 20–30 times more context through these techniques, with month-end close as the example.
Ledgers, aging reports and related documents could be prepared overnight before a finance professional opens the agent.
He calls this ongoing exploration; the description’s blanket “in production” framing is stronger than that spoken status.

Ramp Inspect supplied infrastructure for 10,000 parallel agents examining Ramp’s own systems for vulnerabilities.
A second stage deduplicated reports and attempted adversarial reproduction before escalating findings.
Shevchenko reports seven high-severity vulnerabilities, all patched, including an email-triggered Lua injection chain.
He says open-weight models up to five times cheaper can achieve nearly comparable detection in swarms; no general coverage guarantee follows.

Ramp also applies the work to token-spend analysis, model routing, monitoring and semantic usage categories.
Shevchenko previews transferring post-training work across model versions or families to spread its cost.
Even if a lab exposes a similar feature later, learning how to use it can make adoption faster.

“We need to serve our financial professionals now.” — Alexander Shevchenko, 2:28