Rahul Sengottuvelu is Ramp’s chief technology officer, responsible for the financial-technology company’s technical organization and its application of artificial intelligence. Before assuming that role, he co-founded the customer-support startup Cohere.io, created the open-source structured-generation library Jsonformer, and led applied AI at Ramp.
Sengottuvelu worked on applied AI at Meta and software interfaces at Superhuman before founding Cohere.io with Yunyu Lin and Jason Wang, whom he met at Duke University. Founded in 2020, the Y Combinator-backed company automated customer support using businesses’ existing data; it was separate from the similarly named foundation-model company Cohere. Early experiments with GPT-2, BERT, and GPT-3 required extensive custom infrastructure to compensate for limited context windows and unreliable reasoning.
As Cohere.io’s co-founder and CTO, Sengottuvelu also created Jsonformer, which produces structured data by inserting predictable JSON syntax deterministically while allowing language models to generate the meaningful content. This approach to constrained structured generation makes models responsible for interpretation and conventional software responsible for enforcing the required format.
Agent scaffolding that improves with stronger models. Sengottuvelu favors architectures that become more capable as models and available computation improve. For transaction exports arriving in unfamiliar CSV formats, he compares maintaining provider-specific integrations against giving an agent Python tools, a target schema, and executable tests. Multiple parallel attempts can justify their inference costs when they reduce engineering maintenance and successfully handle previously unseen inputs. His technical explanation of agent scaffolding places verification at the boundary between probabilistic model behavior and deterministic software.
Background coding agents with operational context.Ramp Inspect runs inside isolated development environments with access to repositories, databases, tests, observability systems, and browser-based verification. Its background-agent architecture supports parallel sessions and interfaces spanning Slack, the web, and a Chrome extension.
Risk-calibrated software review. Sengottuvelu argues that code generation shifts bottlenecks toward review and validation. Changes involving authentication, sensitive data, network access, or money movement deserve tighter scrutiny; lower-risk changes can move faster when tests, sandboxing, restricted permissions, and production-like checks establish correctness.
Sengottuvelu has also prototyped an email client in which a language model interprets user actions, retrieves messages, and generates the interface dynamically. Although slow and unreliable, the experiment explores whether future models might execute applications themselves instead of merely helping engineers write them.
An arbitrary CSV import and an experimental mail client expose an architectural choice: should conventional code call the model, or should the model decide when to run code?
