From 13-Second Load Times to Milliseconds: Mercor's Search Rebuilt on MongoDB

In fewer than three years, Mercor has grown from a five-person team in the San Francisco Bay Area to a company valued at $10 billion, increasing revenue roughly 50% month over month, and scaling its head count from a handful of founders to more than 300 employees. Few startups in recent memory have climbed that steep a curve—and fewer still have done it without buckling under their own infrastructure.

Mercor’s AI-native talent network matches candidates to roles using a blend of semantic understanding and precise structured filtering, drawing on resumes, professional profiles, social accounts, and AI-scored video interviews to build a unified candidate view. Behind that network sits MongoDB Atlas, MongoDB Vector Search on Atlas, and Voyage AI embeddings—a consolidated data platform that has grown with Mercor from its days as a member of the MongoDB for Startups program through its current hyperscale era.

THE CHALLENGE

Building an AI-first talent network that could keep up with its own growth

When Pratham Rasal joined Mercor in 2024 as one of its first software engineers, the company was operating on a relational database that was reaching its limits. Rendering a single candidate profile required joining as many as 21 tables, causing the resulting page to take around 13 seconds to load. And traffic was anything but predictable; a single funding announcement could send 1 million users to the platform overnight.

Hiring teams working against tight deadlines need to move through large candidate pools quickly, and every second of load time compounds across every profile they review. “If you are hiring and have to wait several seconds for a single candidate profile to load, it’s a significant drag on productivity,” Rasal said. “You won’t be able to look at more than 100 profiles a day.”

The search pipeline compounded the challenge. Vector embeddings, core application data, and full-text search capabilities were all isolated in separate database systems. Routine changes—such as adding a new candidate attribute, tuning a ranking signal, or adjusting how a field was indexed—required engineers to coordinate updates across three separate services.

When a candidate completed an AI-scored interview, it took roughly 6 hours before they appeared in client search results. If a hiring team needed to fill a role within a tight window, any candidate who finished an interview during that period would surface long after the decision had already been made—costing Mercor placements it was otherwise well positioned to win. For a company whose entire value proposition rests on matching the right person to the right opportunity as fast as possible, the existing architecture couldn’t support the pace of growth the business was experiencing.

OUR SOLUTION

Using MongoDB to power precision matching at scale for Mercor

Rasal came to Mercor with prior experience using MongoDB, recognizing that MongoDB was straightforward to implement and scale as a document database. This made it a natural candidate to benchmark against the company’s existing solution. Rasal and his team evaluated five competing platforms against their internal evaluation criteria, testing each for vector search quality, filtering flexibility, and the operational cost of running another specialized service. One requirement rose above the rest: For a talent network, semantic similarity alone isn’t enough. A candidate in Berlin is not a near-match for a San Francisco–based role, no matter how closely their resume embeds next to the job description.

MongoDB was the best fit as a platform where Mercor could run semantic vector search and hard-filter constraints inside a single query. Location, availability, and credential type became exact filters layered directly onto similarity results, and the rest of the pipeline consolidated onto one system. “In the hiring space, location should be a hard filter,” Rasal said. “It should not be a signal. It should be properly matching or equivalent. That's why MongoDB beats every other service on the market.”

For a team that runs proofs of concept on a near-weekly cadence and ships new features faster than most companies could scope them, the bar for adopting a new solution was high. Any friction in setup, schema design, or iteration would have slowed the startup down in ways it couldn’t afford. But using MongoDB, Mercor could spin up an Atlas cluster, denormalize its candidate data into a single collection, and start testing the new search pipeline against real workloads without the coordinated migrations and specialized infrastructure its previous stack required. Getting started on Atlas was easy—the team was able to bypass traditional sales and contract cycles and self-provision Atlas to move from sign-up to development in minutes.

Today, Mercor runs on MongoDB Atlas with MongoDB Vector Search and Voyage AI embeddings as its unified data layer. Candidate information from every upstream source is denormalized into a single collection, replacing the multi-table joins that used to impede performance. The search pipeline first applies hard filters before combining Vector Search for semantic matching with Lucene-powered full-text search for fuzzy queries, then passes results through Mercor’s ranking algorithms to filter any remaining false positives before they reach a client.

Mercor evaluated more than 10 embedding models, and Voyage AI consistently won on specialized matching tasks where other embeddings fell short. When Voyage AI joined the MongoDB family, two pipelines collapsed into one. “We don’t have to manage two separate services, and we also got early access to Voyage AI releases through MongoDB,” Rasal said. “It was great that we were able to implement these solutions on one platform, and reliability has increased significantly.”

Mercor’s stack also includes Fireworks AI, a longtime MongoDB partner that handles LLM model hosting for the platform. Fireworks AI provides fast, reliable access to the language models that power Mercor’s candidate scoring and profile synthesis, and its tight integration with the broader MongoDB ecosystem means that the small engineering team doesn’t have to manage another vendor relationship to keep the matching engine running.

To keep the matching engine fresh, Mercor replaced its batch process with an event-driven change data capture pipeline built on Apache Kafka, using a managed Kafka service for throughput and reliability. The moment a candidate completes a milestone—such as submitting an interview or updating a profile—a consumer listens for that event and picks up the change. It then generates a new embedding through Voyage AI and writes the result directly to the search collection in MongoDB. Candidates are now searchable the moment they’re ready, rather than hours after the hiring window has opened. Mercor reached a 100% fill rate on time-sensitive client requests as a direct result of this change.

Because MongoDB’s document model doesn’t require migration scripts for schema changes, the team can add new fields as product requirements evolve without downtime or coordination overhead. And when Mercor has asked for capabilities that MongoDB didn’t yet offer, the response has been swift. “We requested a feature for Vector Search, and MongoDB delivered that feature within a month,” Rasal said. “We faced a bottleneck, and MongoDB made sure that we didn’t need to go to another vendor for the solution.”

OUTCOME

Delivering the right match at the speed hiring demands

Since migrating to MongoDB in January 2025, Mercor’s internal search quality score—a proprietary one-to-five measure of how closely results match what clients are actually looking for—has improved threefold. Much of that gain traces back to two changes: the introduction of hard filters and a shift to hybrid search. Because MongoDB stores embeddings and metadata together in a single document, those filtered queries run efficiently without added complexity, making the approach both simple to build and fast to execute. Candidates surface with dramatically higher relevance, so team members spend less time filtering out poor matches and more time evaluating the people they actually want to hire.

Profile page latency dropped from roughly 13 seconds to milliseconds. Team members who used to spend a full workday reviewing 100 candidates can now move through the same pool in minutes. “If you’re getting a profile in milliseconds, you’ll be able to search more, you’ll be able to retrieve more, and you’ll be able to hire faster,” Rasal said. “That’s a major difference that we’ve seen from our clients.”

The underlying database has grown with the business. It launched on MongoDB at 100 GB and has since crossed 2 TB as of May 2026, scaling through four cluster tier upgrades from M30 to M200. Auto-scaling handles the traffic spikes that Mercor experiences whenever a product launch, funding announcement, or press moment drives a sudden wave of users to the platform. No one has to sit in front of a dashboard at midnight when a launch goes live; instead, the team can ship with confidence that the infrastructure will absorb whatever traffic follows. And because new candidates now become searchable within moments of completing an interview rather than hours later, Mercor’s sourcing team can surface talent at the speed its clients are ready to hire.

That foundation has empowered a very small engineering team to build at a pace most companies can’t match. “Using really good software tools like MongoDB and being able to use effective LLM tooling enables us to write stable systems faster,” said Scott Bruce, Principal Engineer at Mercor. “This means we can be incredibly effective with a very small number of people.”

This rapid pace and relentless agility are part of what draws engineering talent to the company. Engineers at Mercor work across the stack on problems that often ship the same week they’re scoped, providing the range and ownership that’s hard to find at more established companies. “Mercor is a very good blend of extreme hypergrowth and extreme commitment to making sure things get shipped, that they’re useful to the world, and that we do so very quickly,” Bruce said. “It’s a great engineering culture if you care about setting up foundational pieces that help the company grow and exist better.”

The impact of that speed extends well beyond Mercor’s own team. The network has connected a massive talent pool to U.S.-based startups and AI companies that might never have found them otherwise, and the income candidates earn through Mercor has covered rent, funded relocations, and reshaped livelihoods. Traditional hiring platforms are effective at listing open roles, but matching candidates to work that genuinely fits their skills and circumstances has remained a persistent gap. By closing that gap at scale, Mercor is changing the economic trajectory of the professionals on its platform.

“With the AI evolution, everything is changing so fast that you have to keep up with the current technology,” Rasal said. “You have to make sure that the thing you are making is the best. If you’re not ahead, if you're not selecting a solution that can scale very well, if you’re still stuck with a legacy system, you are not providing the best experience to your customers.”

Looking ahead, Mercor is pushing deeper into index tuning and vector quantization to sharpen search performance as data volumes keep climbing. The team is also exploring hybrid retrieval workflows across its newer verticals in enterprise, agentic workflows, and benchmarking, including its recently launched APEX-SWE benchmark for software-engineering agents.

Mercor reevaluates its stack on a regular cadence, and MongoDB keeps earning its place. “We are still sticking with MongoDB because it is evolving with the technologies,” Rasal said. “It’s providing all the solutions we need.”