Biography

About Ravi Singhal

Ravi Singhal is a Principal Engineer (Executive Director) at Wells Fargo, where he builds agentic AI platforms and the LLM inference systems underneath them. He has been an engineer for twenty years and still writes production Go, Python and Java. He lives in Gurgaon, India.

Wells Fargo

At Wells Fargo he owns the agent platform, the tool plane, agent identity and the evaluation layer behind three programmes: commercial banking AI, API monetization and consumer open banking. Nine teams and around a hundred engineers build on that platform. Architecture is set jointly with engineering, model risk and product.

He wrote the bank's engineering standard for validating non-deterministic, tool-using agents, covering reproducible replay, trajectory-level assertions, model inventory registration and independent validation. He moved the commercial banking assistant off a decentralised swarm onto a supervisor topology, so every agent handoff is logged as a replayable decision record and anything irreversible stops for a human. He designed how the tool plane authenticates and authorises access to treasury and payments data, built on SPIFFE workload identity and delegated token exchange, with policy checked at the gateway on every tool call.

ServiceNow

Before Wells Fargo he was a Senior Staff Software Engineer on ServiceNow's Enterprise AI Platform. He defined the multi-agent reference architecture on AWS EKS and Bedrock, using LangGraph for stateful orchestration, the Model Context Protocol for tool access, bounded autonomy and Kafka-driven event streaming. He architected a multi-provider model gateway across OpenAI, Anthropic and Bedrock-hosted models, routing on cost, latency and capability, and cut inference spend by 30% while holding the p95 latency targets.

The work he describes most readily is an evaluation failure. A silent fallback was scoring failed model calls as passes. He rebuilt the harness to fail closed, and reported accuracy corrected from 91% to 71%. That lower number is what the team shipped against.

Oracle and Verrazzano

Ravi spent nine years at Oracle as a Consulting Member of Technical Staff (Senior Principal) and Developer Advocate. He was thefounding architect of Oracle Verrazzano, the open-source multi-cluster Kubernetes platform that reached general availability in 2021. He designed its multi-cluster architecture, built the Go operator-driven control plane, and shipped Kubernetes operators that automated a nine-component stack through declarative custom resources. It grew to more than 200 tenants, ten internal Oracle teams and five external enterprises. He closed its first enterprise deal, worth $1M with a large Japanese enterprise, working the sales cycle forward-deployed as the solution architect. He built and led the twenty-engineer platform team behind it, owning hiring and technical direction.

He shipped two further products into Oracle's cloud lineup: Wercker, a container-native CI/CD service, and OCI CloudShell, the browser-based shell that ships by default in every Oracle Cloud tenancy. From 2023 he architected a production generative AI stack on Kubernetes with vLLM serving, agentic retrieval over Oracle Database, and evaluation gates using RAGAS and LLM-as-judge scoring. He published nine technical articles across five language ecosystems.

Banking and capital markets

At NatWest Group, then RBS, he was a Technical Specialist across RBS Digital and UK Banking HQ. He owned the prototype and evaluation of the bank's first streaming machine learning pipeline, on Kafka, Storm and LingPipe, and the recommendation that came out of it. He architected a zero-downtime omnichannel platform for more than 20 million users on Java Spring Boot microservices.

Earlier, at Publicis Sapient, he spent four years onsite inside UBS Wealth Management and Wellington Management, owning the data infrastructure behind quantitative research and pricing. He built distributed Hadoop pipelines for quantitative parameter computation and a real-time options pricing engine over Thrift-based distributed RPC. His career began in 2006 in telecom OSS/BSS and enterprise BPM tooling, at Amdocs, Cordys (now OpenText) and HCL.

Open source

He maintains two open-source platforms that carry his current thinking in runnable form:llm-inference-platform, a Kubernetes control plane for LLM inference that autoscales on queue depth rather than GPU utilisation, andagentic-adjudication-platform, an adjudication system that can grant but can never refuse, because the two errors do not cost the same.

Education and contact

Ravi holds a B.Tech (Hons) in Computer Science and Engineering from Maharshi Dayanand University, Haryana. He can be reached at ravi@ravisinghal.io, on LinkedIn, on GitHub, and on X.