CV Deep Dive

Today, we’re talking with Sharath Rajasekar, Founder and CEO of Javelin.

Javelin is an enterprise AI security platform designed to serve as a real-time enforcement layer between internal applications and foundation models. Founded by Sharath and his co-founder Anil— former operators at Oracle and Walmart Labs—Javelin, first of its kind enterprise grade closed-loop security platform that provides low-latency LLM gateways, security-focused small language models, and agentic tooling to help companies secure, govern and scale their AI workflows from development through production. Their core product helps detect and block threats like prompt injection, jailbreaks, tool misuse, and data leakage, and integrates directly into existing enterprise infrastructure with minimal friction.

Today, Javelin is being used across Fortune 500s and AI-forward teams that are moving real workloads into production. Whether it’s securing AI agents, protecting proprietary data from leakage, or automating compliance and risk review across dozens of LLM-powered tools, Javelin sits directly in the path of model communication to enforce policy and detect threats in real time. Its low-latency gateway architecture, enterprise-grade deployment options, and suite of security-focused language models make it a foundational layer for organizations serious about adopting AI safely at scale.

In this conversation, Sharath shares why AI security needs its own stack, how Javelin built a high-scale platform in Rust and Go, and why small, specialized models—not general-purpose ones—will define the next wave of AI security.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Sharath 💬

Sharath, welcome to Cerebral Valley! First off, introduce yourself and give us a bit of background on you and Javelin. What led you to co-found Javelin?

Hey there! My name is Sharath and I’m the co-founder and CEO of Javelin. Before Javelin, I spent many years at Oracle—most recently running Oracle’s Customer Data Platform, which is a petabyte-scale data platform. Before that, I spent about a decade working in voice over IP and WebRTC, focusing on security and the integration challenges of embedding web audio and video into applications. That’s a big part of my background.

My co-founder Anil comes from a deep enterprise background. He was formerly a Distinguished Architect at Walmart Labs, where he helped build out Walmart’s big data and e-commerce infrastructure. So both of us come from building and scaling large, petabyte-scale cloud systems globally—and our goal was to bring that same level of rigor to building Javelin and its product stack for enterprise AI adoption.

With the advent of AI, one of the core challenges my co-founder and I noticed was around adoption. There’s a huge diversity in models now, and that’s only accelerating. But along with that proliferation comes a set of risks—controls, policies, and security threats that are not easily handled in a multi-provider, multi-modal context. The absence of tooling to deeply understand model inputs and outputs or mitigate data risks has made it harder for enterprises to adopt AI safely and confidently at scale. That’s the problem we wanted to solve.

How would you describe Javelin to the uninitiated developer or AI team?

One of the biggest impacts the internet had on software was making it easy for companies to operate, do business, and communicate electronically. But what we often overlook is that it was the underlying networking and security fabric—things like SSL and other security protocols—that made that all possible. That’s what enabled cloud computing to take off.

We see ourselves doing something similar for AI. For AI to be widely adopted and trusted, there has to be a strong underlying security layer. That’s the fabric we’re building at Javelin—something foundational that allows enterprises to safely scale their AI workloads and workflows.

Javelin is essentially an AI security platform - a protective layer that sits between your application—or the agents you're building—and both the tools, MCPs and the AI models you're using. Being in that path, we inspect all the AI and agentic traffic that flows through the system.

We can deeply inspect and filter incoming prompts and model responses for security. Think security guardrails—everything you’d want to apply from a security posture standpoint: malware detection, phishing detection, prompt injection, jailbreak attempts, content moderation, safety checks, and more. Model & agentic security is a fast-evolving category, and the set of protections is constantly expanding and Enterprises rely on us to protect their data and applications.

Who are your users today? Who is finding the most value in what you're building at Javelin?

Almost every organization adopting AI is thinking about security in some form. Whether you're building AI-native apps, deploying agents, or just enhancing existing workflows, it's easy to get something working in a sandbox or proof of concept. But when you move toward production, that’s when security and legal reviews start to slow things down.

Having Javelin in place early on gives teams confidence that they're already aligned with key organizational security standards. The teams we typically work with are mid-to-large AI forward enterprises—including a number of Fortune 500s.

The groups that see the most value right away are usually the security teams: SecOps, CISOs, and those responsible for overall application security, but CTOs and software teams also benefit significantly. Javelin serves as an AI-aware egress layer that’s designed to be as transparent as possible. That means all your agentic applications can route AI traffic —while still establishing a solid perimeter of AI-specific security. Thanks to Javelin’s distributed architecture, software teams gain scalable tools that help them manage and secure AI adoption across large, complex environments.

Which existing use-case for Javelin has worked best? Any customer success stories you’d like to share?

Our use cases really span the whole ecosystem of AI applications being built today. In One of our largest deployments, a global enterprise runs dozens of AI-powered applications through a federated, multi-region Javelin deployment. Our platform provides them both a central “pane of glass” for enterprise-wide security policy management and a fully federated control plane for individual teams. Beneath that enterprise layer, each product or line-of-business team gets its own isolated “sub-space” where they can tailor settings to their needs—selecting which small models to use, adjusting risk-scoring thresholds, enabling or disabling specific tool integrations, and granting access only to approved roles.

Because our control plane is multi-region and multi-tenant, you can spin up new AI workloads in any geography or business unit in minutes, confident that the same centralized policies will be enforced everywhere. The result is a security posture that’s both consistent across the enterprise and flexible enough to support each team’s unique AI initiatives. That also means every application and agent automatically benefits from real-time enforcement of security controls — all without changing a single line of code.

Another example, a customer uses Javelin both in development and production. During development, our agentic testing tools act like adversarial users—scanning for OWASP Top 10 and MITRE-style vulnerabilities before code ever ships. In production, that same intel drives our platform’s runtime enforcement: detected weaknesses become active guardrails that block attacks in real time. This dev-to-prod feedback loop is critical to robust AI security.

Walk us through Javelin's platform. What use-case should users experiment with first, and how easy is it for them to get started?

If you’re on an application team, you can connect your CI/CD to Javelin RED—our adversarial, agentic red-teaming toolkit—and instantly scan your app to see how it holds up under real-world attack vectors right within your software development lifecycle.

If you are a security leader, we offer multiple deployment paths depending on the scale and complexity of the company. It’s super easy to get started—Javelin can be hosted by us in a fully managed SaaS model, or deployed in your own cloud. If you're running in AWS or GCP, we support bring-your-own-cloud setups with automated deployment. We can drop Javelin directly into your VPC, and with a single line code change, route all your app traffic through it.

From that point on, there’s nothing else to configure. All your applications automatically get the benefits of Javelin’s security capabilities—guardrails, content filtering, real-time enforcement—right out of the box.

How are you measuring the impact or results that you’re creating for your customers?

There are lots of metrics that we look at. By combining raw threat-detection numbers with real-world model alignment, we give customers both the tactical insights - ‘what went wrong’ and the strategic assurance - ‘hey, you are secure per current standards’. One metric our customers really find value in is how many AI threats we are able to highlight or escalate to their security. Whether it's a prompt injection that was detected, a security guardrail that was triggered and a prompt rejected, or some other security violation that was introduced into the workflow without them realizing—it’s those sorts of things that really matter. Those are the events that directly translate into avoided breaches, reduced risk, and greater trust in production AI.

Beyond threat counts, we also: validate production readiness, show application teams exactly how their AI holds up against real-world attack vectors, align with emerging AI security standards and map our findings to frameworks like the NIST AI Risk Management Framework, OWASP, MITRE and ISO JTC guidelines—giving customers evidence they can share with auditors and executives.

There are a number of companies working on AI. What sets Javelin apart from a user/product or technical perspective?

When we founded Javelin, our mission was simple: build the first truly enterprise-grade AI security platform. What we’ve seen in the market are really point solutions—tools that solve one narrow problem but don’t integrate into a broader AI security ecosystem. They check a box, but leave gaping holes in your end-to-end protection.

Javelin is different because we deliver closed-loop security across the entire AI lifecycle: We scan your applications and models for vulnerabilities before you ship and prioritize real-world threats and generate updated guardrails. Our ultra-low-latency AI Gateway ingests those guardrails and dynamically blocks attacks in production.

Pair that with our growing set of research backed security guardrails—built using a combination of traditional cybersecurity tooling and small language models—and we’re in a really strong position. On top of that, we’ve now built an agentic toolkit for testing applications and closing the loop between dev and production. It’s that combination—low-latency infrastructure, robust security primitives, and closed-loop tooling—that we think makes us truly unique.

That seamless feedback loop—where dev-time findings instantly inform prod-time defenses—is unique in the market. Instead of stitching together disparate point tools, you get one unified platform that learns, adapts, and secures your AI from code check-in through user interaction. That closed-loop model is something we think is pretty unique.

Could you share a little bit about how Javelin actually works under the hood?

We’ve always been obsessed with scale. From day one, we chose Rust and Go to build our low-latency LLM platform —while most open-source gateways rely on Python or JavaScript and buckle under heavy load. As AI workloads spike and context lengths surpass millions of tokens those solutions start to fall apart. Our engineering team poured months of work into a transparent egress layer that can handle billions of AI transactions with sub-5-millisecond latency. That foundation isn’t just a technical boast; it’s what keeps mission-critical AI applications running smoothly when everything else slows to a crawl. That foundational layer has been critical, and I think it’s something that sets us apart.

In fact, we have a couple of research papers coming out soon around novel transformer architectures we’re working on, specifically tailored for security-focused models. Model selection is a really interesting space because you’re always having to find the right tradeoff between accuracy, speed, and cost. That’s true even for the models we build internally.

We’ve been focused on creating very small language models—usually around a billion parameters or less—that are highly specialized to solve security-related problems. These models are designed to respond extremely quickly, because latency in the signal path is critical.

It’s taken a lot of iteration to overcome those challenges—choosing the right transformer architecture, the right embedding strategy, the right training pipelines. Security is a tough domain because the datasets are incredibly hard to come by, and the benchmarks for evaluating model performance are just as scarce. So we’ve had to build all of that from the ground up—not just the models, but the datasets, the benchmarks, and the entire training pipeline. We’re starting to open-source some of that now as well.

Given the excitement around new trends in AI such as Agents and Multimodal AI, how does this factor into your product vision for Javelin?

That’s a great question. At Javelin, we’re building the security fabric for the next generation of AI—agentic and multimodal. AI agents are top of mind for almost everyone right now. We’re seeing a shift where software—everything from workflows to systems that have been built over the past 15 to 20 years—is starting to be disrupted by the move toward agentic frameworks. It’s clear this future is coming.

Now, agents still have a way to go in terms of maturity—both at the foundational model level and in terms of the security you need to wrap around them. We’ve been focusing on agentic workflows and how to bring security into that context. In fact, we’re already doing early work with some of the agentic frameworks to build what you might think of as security in the loop.

Building security for agents and applications that deal with very-large context windows or multi-modal models is also frontier research and something we are very interested in.

How do you see Javelin evolving over the next 6-12 months? Any specific developments that your users/customers should be excited about? One area we’re tracking closely is the rise of agents, tool calling, and model context protocol (MCPs) usage —which are quickly becoming core to how agents operate. With that comes a new set of challenges: tool poisoning, excessive agency, indirect prompt injections and others. We’re focused on solving those problems and will be investing heavily in them over the next 12–24 months.

Looking ahead to 2025, our small language models will continue to evolve, and we’re excited to expand into multimodal and multilingual use cases. As the threat landscape grows, we’re growing our portfolio of specialized LLMs accordingly.

How would you describe the culture at Javelin? Are you hiring, and what do you look for in prospective team members joining Javelin?

What makes Javelin special is that everyone on the team is driven by first-principles thinking. We’re in a moment where everything in the industry is changing, and it requires rethinking the way we build, secure, and deploy software from the ground up. We move fast, we operate with a high degree of ownership, and we’re true builders—innovators who care about the details.

We’re also a distributed, remote-first organization. We have team members across North America and engineers all over the world working on Javelin, and we’re excited about the diversity of perspective and velocity that it brings. We are hiring!

Job Openings:

Software Engineer - https://www.linkedin.com/jobs/view/4217432261/

Software Architect - https://www.linkedin.com/jobs/view/4217427720/

Conclusion

Stay up to date on the latest with Javelin, learn more about them here.

Read our past few Deep Dives below:

• TensorStax is Driving The Future of Data Engineering 🎛

• Zencoder’s Repo Grokking - Your AI coding unlock 🔑

• Rime - Authentic AI Voices for Real-time Conversations at Scale 🗣️

• Future AGI - Built for End-to-End AI Development 💡

• Antithesis - the last word in autonomous software testing 🎛

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email (newsletter@cerebralvalley.ai) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs