Keewano Raises $12M for AI-Native Database Infrastructure

Keewano has raised a $12M Seed round to launch KeewanoDB, a database built around a customer that does not wait for a data team to model the next question: an AI agent. The company argues that agents need to investigate large event histories, change direction as they learn, and keep the surrounding context intact while they move.

Hetz Ventures led the financing. a16z speedrun, Remagine Ventures, and DIG Ventures participated alongside unnamed angels. The round was announced on September 15, 2026, as Keewano came out of stealth with the first public release of its database.

The transaction does not disclose a valuation, security type, individual check sizes, named angels, board changes, prior funding, cumulative capital, or a detailed use-of-proceeds plan. What it does reveal is the technical bet behind the capital: databases designed for predefined human queries may not fit the way autonomous software explores data.

The Seed Round Backs a Different Query Pattern

Traditional analytics systems are usually prepared around known questions. A team chooses the metrics, organizes data for those questions, and builds dashboards or queries that can be reused. An agent works differently. It can start with one question, discover an unexpected pattern, and issue another query that no one planned when the data model was created.

Keewano says KeewanoDB keeps each entity's sequence of events together and in order. It then runs analysis inside the database, aiming to reduce the repeated preparation required when an agent asks a new question. The product thesis is less about making a familiar dashboard faster and more about preserving enough context for software to investigate without reconstructing the story around every event.

That distinction matters if AI agents become regular consumers of operational data. A fraud agent, security system, game-operations agent, or customer-support agent may need to follow behavior across time rather than retrieve one isolated row. A fast lookup can still produce a weak conclusion when the sequence before and after the event has been broken apart.

Keewano and Hetz publish performance claims for the product, but the figures vary across current company and investor materials and have not been independently audited. The financing announcement therefore offers a credible product direction, not a settled benchmark comparison. Production deployments, reproducible tests, and customer outcomes will have to supply that evidence.

A Team Built Around Data-Intensive Products

Keewano's current leadership brings together four repeat founders and operators. Mark Kardashov is CEO, Dima Karger is COO, Pavel Bibergal is CTO, and Vitaly Bukhovsky is CISO. Current role information comes from the live a16z speedrun company profile; older directory pages carry stale title assignments.

Kardashov and Bukhovsky previously built TestProject, a test-automation platform acquired by Tricentis in 2019. Bibergal served as CTO at Plarium, while Karger ran one of the game developer's largest studios. Hetz also points to Devalore, acquired by abra, as part of the team's prior company-building record.

That history is relevant because Keewano's original problem emerged from game operations. In an official founder account, Bibergal describes analysts trying to understand player behavior across enormous event streams. The challenge was not simply storing another event. It was retaining a player's full behavioral sequence so the system could examine what happened before and after a choice.

The company first applied that architecture to gaming analytics, then broadened it into infrastructure for AI agents. The product category changed, but the underlying demand stayed consistent: preserve the sequence, run the investigation close to the data, and avoid rebuilding context for every new line of inquiry.

Why Hetz Led the Investment

Hetz frames Keewano as both a technical architecture bet and a team bet. The lead investor's announcement emphasizes that the founders have shipped data-intensive products, led large teams, and completed prior acquisitions. It also identifies machine-generated investigation as a workload existing databases were not originally designed to serve.

The participant group extends that thesis across AI infrastructure and early-stage software. a16z speedrun lists Keewano in enterprise software, AI agents, and AI infrastructure. Remagine invests in technology shaped by changes in media, data, and consumer behavior, while DIG backs early-stage software founders. None of the investors disclosed individual check sizes or ownership stakes.

The round arrives before Keewano has published the commercial evidence that would settle its category claim. The announcement does not name customers, revenue, retention, production volumes, or a deployment pipeline. That makes the capital useful for product and market development, but it does not by itself prove that buyers will replace or supplement established data platforms.

What KeewanoDB Now Has to Prove

KeewanoDB has to show that its context-preserving model improves actual agent decisions, not only query speed. That will require workloads customers can reproduce, clear comparisons against existing database and analytics stacks, and evidence that the system remains reliable as data volume and query unpredictability increase.

Integration will matter as much as raw performance. Enterprises already operate warehouses, lakehouses, streaming systems, governance layers, and security controls. Keewano can become useful without replacing all of them, but customers need to understand where it sits, which data moves into it, how access is controlled, and what operational burden follows the deployment.

Developer adoption creates another test. A database for agents still needs people to evaluate it, connect it, observe it, and trust its behavior. Documentation, tooling, ecosystem support, and transparent benchmarks will decide whether the architecture travels beyond the founding team's first set of use cases.

The $12M Seed gives Keewano time to build that proof while AI teams are still deciding what their agent data layer should look like. Its strongest idea is already clear: an agent cannot reason from context that the database discarded. The next evidence will come from the companies willing to put that idea behind a live agent, let it ask the question nobody modeled, and measure what arrives with the answer.