Building a voice AI interviewer that can retrieve candidate context in real time
Casuro is building a new approach to technical hiring with VOCAL, a voice AI agent that sits directly on a company's job or career page and has conversations with candidates as part of the application process. Rather than relying solely on resumes and traditional assessments, VOCAL gives candidates a way to talk through their experience, demonstrate what they know, and answer questions that adapt based on the context of their application.
The idea came from Naman Shukla, founder and CEO of Casuro, while he was hiring data scientists and ML engineers himself. He was looking for alternatives to platforms like LeetCode and HackerRank that would give candidates a better way to demonstrate their skills, but couldn't find anything that solved the problem.
“We started scouting and looking for any alternatives to LeetCode or HackerRank kind of assessment platforms, and we couldn't find any that would actually give somebody an ability to demonstrate their skills.”
As AI became more widely used in the hiring process, the problem became even more apparent. Candidates could use tools like Claude and Codex to tailor their resumes and applications, making it harder for hiring teams to understand what someone actually knew or how they thought. Casuro's approach was to add more context to the process rather than trying to remove AI from it.
VOCAL can have a conversation with candidates directly on a company's job page, answer questions about the role, understand their experience, and ask increasingly relevant questions based on what it learns. That context then becomes available to the hiring team, giving them a richer picture of the candidate than a resume or application form alone.
I sat down with Naman for our Built with Moss series to talk about how they built VOCAL and what it took to move the product from an early prototype into production.
From a voice demo to a production system
The idea behind VOCAL is straightforward. The infrastructure required to make it work reliably at scale is not. During a conversation, the agent needs to understand what a candidate is asking, retrieve relevant information about the job and candidate, and respond naturally, while Casuro continuously indexes candidate information as applications come in.
The team initially worked with different parts of the voice AI stack, including Pipecat and LiveKit, but as they started thinking about real production usage, the limitations of their first version became obvious.
“The first version broke or slowed down at three to five concurrent conversations.”
That forced the team to think beyond getting a voice agent working in a demo. What would happen when 100 or 1,000 conversations were happening at the same time? Could the system keep indexing candidate information while conversations were happening? And could the agent retrieve the right context quickly enough that the conversation still felt natural?
For Casuro, these were all part of the same product problem. A voice interviewer needs to be able to handle growing concurrency while continuing to retrieve the right candidate and job context without introducing delays into the conversation.
“Imagine like 100, 1000s of concurrent conversations going on. It was super important for us that the RAG works, the scalability should not break anything, and more importantly, it should sound human.”
Why retrieval became critical
One of the biggest technical challenges was making candidate and resume information immediately available to VOCAL during a live conversation. Casuro needed to pull together information from applications and recruiting systems, including ATS platforms such as Ashby, and continuously index that context so the agent could use it when interacting with a candidate.
This made retrieval a core part of the voice experience. In a traditional application, a few hundred milliseconds of additional latency might go unnoticed. In a voice conversation, even a small delay can make an interaction feel less natural.
This is where Moss became an important part of the architecture. Casuro uses Moss as the retrieval layer behind VOCAL, giving the agent fast access to candidate and job context while allowing the Casuro team to focus its engineering effort on the voice experience and hiring workflow rather than building and maintaining the underlying retrieval infrastructure.
“With Moss, their RAG system is under 10 millisecond retrieval, which is awesome.”
That performance gives VOCAL a fast path to the context it needs while a candidate is speaking with the agent. During our demo, the agent recognized that Naman was a founder who had personally onboarded customers and used that context to ask him a targeted question about owning post-sale relationships.
“It's hyper tuned questions to get more and more information about you.”
That's the experience Casuro is trying to create: not another form or static assessment, but a conversation that can adapt based on what it learns.
Turning better context into better hiring decisions
Once the conversation is complete, Casuro makes the resulting context available to the hiring team. Recruiters can access transcripts, summaries, and information gathered during the interaction, while Casuro's integrations bring that context into existing recruiting workflows. The company supports ATS platforms including Greenhouse, Ashby, and Lever, and has also built MCP integrations that allow teams to work with candidate information through AI tools.
The product itself has evolved based on what customers actually want from the process. Some hiring teams want AI to evaluate candidates against a rubric, while others simply want to gather more context and leave the evaluation to a human.
“We designed the solution along with them.”
That flexibility has become an important part of Casuro's approach. The goal isn't necessarily to automate every part of hiring; it's to help teams gather better information earlier, then decide how much automation makes sense for their process.
Naman shared the example of Gifts AI, a small team that was hiring for a founding engineer. After posting the role on LinkedIn, the company received so many applications that it had to stop accepting them after roughly half an hour. Casuro helped the team create a different process, asking candidates to demonstrate what they could actually do and pitch themselves to the hiring team rather than relying entirely on a traditional application.
Gifts AI ultimately hired its founding engineer through Casuro.
“They hired their founding engineer entirely through us, from zero to actual hire.”
Casuro has also seen the conversational application experience help reduce automated applications. Naman said the approach has reduced bot applications by roughly 40%.
Choosing infrastructure that can scale with the product
One of the things that stood out from our conversation was how quickly infrastructure decisions became product decisions for Casuro. The candidate doesn't know what retrieval system is powering VOCAL. They just know whether the agent responds quickly, whether the conversation feels natural, and whether the follow-up questions actually reflect what they said.
As Casuro moves toward supporting more candidates and more concurrent conversations, the underlying infrastructure needs to scale without becoming something the team constantly has to manage. That includes the retrieval layer responsible for indexing and surfacing candidate and job context fast enough to keep the voice interaction moving.
Naman put it this way:
“In a voice conversation, stability and latency is the product. A candidate doesn't know or care what retrieval system sits behind VOCAL. They know whether the agent paused before answering and whether the follow-up question showed it actually understood them. Moss gave us sub-10 millisecond retrieval on candidate and job context, and that's the difference between a conversation that feels human and one that feels like a form with a microphone.”
For Naman, the infrastructure decision is ultimately about making sure the product can scale without forcing the team to constantly revisit foundational systems underneath it.
“It's super important to choose the right infrastructure at the beginning, so that once you scale, it's very easy for your agents to scale.”
For Casuro, Moss became that retrieval layer. The team can continue building the voice and hiring experience while relying on Moss for the fast retrieval and indexing capabilities underneath it, allowing the infrastructure to scale alongside the product rather than becoming a bottleneck as concurrency grows.
That's increasingly true across real-time AI applications. As agents become more contextual and interactive, retrieval isn't just a backend concern. It becomes part of the product experience itself.
For Casuro, the goal is to give candidates a better way to show what they can do and give hiring teams more context before they make a decision. Moss helps make that experience possible.
Built with Moss
Casuro is building VOCAL to give companies and candidates a more contextual, AI-native approach to technical hiring.
Moss powers the retrieval layer behind VOCAL, providing fast access to candidate and job context while giving the Casuro team the infrastructure to scale toward more concurrent voice conversations.
Building an AI application that needs fast, scalable retrieval? Explore Moss
