Startup Spotlight: Upsolve AI builds analytics agents that have to show their work
Upsolve AI's Ka Ling Wu and Serguei Balanovich carried lessons from Palantir's HyperAuto into a product that gives analytics agents company definitions, permissions and query checks.
By Ryan Merket · Published · Updated
Primary source: Y Combinator
Why it matters
Upsolve's founders are productizing enterprise data work they previously did at Palantir. The test is whether governed context and evaluation can make AI analytics repeatable across customers without bespoke implementation.
Ka Ling Wu (@wukaling) and Serguei Balanovich are building Upsolve AI around a problem they met at Palantir: companies can connect software to data and still fail to make that data useful to the people who need it. Its product, Agent Studio, is designed to give analytics agents the business definitions, access rules and query history needed to answer questions about a company's own numbers. The founders' bet is that those controls can turn a convincing demo into a tool a business can actually use.
The Y Combinator profile lists Upsolve in the Winter 2024 batch and describes the product as a platform for building, testing and monitoring data agents. The profile does not announce a new launch. Before starting Upsolve, Wu and Balanovich helped build Palantir's HyperAuto, which the founders say grew to more than 50 enterprise customers and eight figures in annual revenue.
A product built from the founders' last job
Wu spent four years at Palantir, where she worked across client engagements and led HyperAuto from its creation through launch. She later led product-solution development and client expansion at nPlan, an AI company focused on construction technology. Balanovich spent seven years at Palantir, working on supply-chain and manufacturing deployments, productizing that work into HyperAuto, and serving as a development lead on Palantir Apollo, according to their YC profiles. Before starting Upsolve, the pair had experience turning difficult data work into products used by enterprise customers.
Wu has described the decision to start Upsolve as a search for a problem, rather than a commitment to one idea. In an interview about the company's origin, she said the founders initially considered leaving data behind for ideas such as a travel or language-learning app. Their previous work kept drawing them back. Wu said they saw problems from their Palantir experience that AI might automate, and wanted to apply those lessons to companies smaller than Palantir's traditional enterprise customers.
The first version of the idea focused on internal analytics. Conversations with startup founders changed the target: those companies wanted analytics for their own users, too. Upsolve initially concentrated on embedded analytics, giving software companies a way to offer dashboards and data exploration inside their products. Its current positioning also reaches internal data teams, which can deploy agents for employees and connect them to tools such as Slack, Teams and other work surfaces.
Upsolve serves two related buyers. A data or BI team may want fewer repetitive analyst requests; a software company may want to give customers useful analytics without building and maintaining a reporting stack itself. Both need to make data usable while keeping metric definitions consistent and restricting access to information users should not see.
The answer needs to show its work
Upsolve's Agent Studio is built around assembling that context. The product connects to warehouse tables and existing dbt projects, then lets teams organize definitions, business rules and institutional knowledge alongside validated SQL patterns. Its product materials describe tracing an agent's path from a user question through generated SQL and tool calls to its final answer. The interface can show the metric definition or query pattern used, and the company says teams can test and monitor agents against benchmark questions.
That approach addresses a familiar failure in natural-language analytics. A query can be syntactically valid and still produce the wrong business answer. One department may count refunds differently from another; a user may have permission to see an aggregate but not individual customer records. A language model cannot resolve those distinctions just by reading the warehouse schema. Someone has to define the business meaning, permissions and checks, then keep them current.
Upsolve's product materials divide the job into data structure, business meaning and trust. Agent Studio is designed to let data teams inspect how an answer was produced, including its context and query path. The company also says it supports row-level security, end-to-end tracing, evaluation and monitoring. These are product features and design claims; they do not, on their own, establish how accurate an agent will be on a particular company's data.
The company's own pages give different totals for supported SQL connections. Its product page says 30-plus, while its pricing page says 50-plus. Both pages name common warehouses including Snowflake, BigQuery, Redshift, Postgres and Databricks. The counts describe the company's offering, not independent measures of performance.
From customer-facing dashboards to agents
Upsolve first focused on embedded dashboards and analytics inside other software companies' products. Its newer Agent Studio framing emphasizes how data teams prepare context and govern answers before they reach end users, in line with the founders' stated aim of making data more directly useful in decisions.
Customer feedback on Upsolve's website points to the embedded use case. Fiber AI's co-founder says the product helped the company customize campaign and marketing insights for customers, while PAXAFE's co-founder describes using it to deploy tailored supply-chain dashboards. These are customer testimonials published by Upsolve, not independently measured case studies. They illustrate the problem the company says it serves: a software vendor needs analytics that can reflect customer-specific data and permissions without asking its own engineering team to build every dashboard from scratch.
The internal-agent use case has a larger competitive field. Existing analytics vendors are adding natural-language and agent features to products that already include data models and governance. ThoughtSpot's March 12th, 2026 product announcement describes semantic context for analytics. Meanwhile, Credible Data announced a $10 million seed round in July 2026 for a platform focused on governed business context that AI systems can reuse. These products differ in scope and approach, but compete for the same resource: a reliable account of what a company's data means.
Upsolve's prospective advantage is the combination of context management, agent testing, permissions and deployment across internal and customer-facing surfaces. Its challenge is proving those pieces are easier to operate together than the tools a company already owns. For buyers, the key test is whether analysts can trust answers, correct bad ones and prevent access-control mistakes without adding manual work.
What the early numbers show
The clearest operating figures in the founders' account come from their previous product, not Upsolve. The YC listing says HyperAuto reached more than 50 enterprise customers and eight figures of annual revenue in two years. Upsolve's About page describes the same growth as more than 50 customers in less than 2.5 years. Both are company accounts of HyperAuto's history; they are not Upsolve customer or revenue figures.
For Upsolve itself, the YC profile says the product is in production with Fortune 500 users, BI teams of more than 100 people and growth-stage companies. The statement identifies the kinds of organizations the company targets, but does not give a customer count or measure use, retention or revenue. The company's website also features named customer testimonials, while leaving the commercial scale of those examples unclear in the material reviewed here.
Upsolve secured outside backing in 2024. On September 18th, 2024, investor PS27 Ventures reported that the company won a $1 million investment from HubSpot Ventures through HubSpot's Million Dollar Pitch competition. SmartCompany reported the prize as $1.5 million in Australian dollars and said the winner also received $25,000 in AWS credits. The figures describe the same competition prize in different currencies, not separate rounds. Wu told SmartCompany that the company planned to use the money to hire engineers and build the product. The financing milestone is two years old; it does not establish Upsolve's current funding total or valuation.
The current pricing page gives a view of how the product is packaged. Upsolve offers a free tier, a Pro plan listed at $500 per month, a Team plan at $2,000 per month, and custom Enterprise pricing. The paid plans allocate monthly credits, with row-level security, embedding and multi-tenant support listed in the Team tier. Enterprise includes options such as private deployment and bring-your-own-model support. The prices show the entry point, but not how many customers pay them or how much support an enterprise deployment requires.
Upsolve's enterprise offer includes Forward Deployed Engineering, with company staff helping customers build data models and encode internal rules. That can help a customer get past the hardest early work, especially when definitions and permissions are scattered across teams. The product's long-term efficiency depends on how much of that setup can become repeatable software rather than bespoke implementation.
Wu and Balanovich have seen the problem at enterprise scale. Upsolve's test is whether data teams can maintain definitions, access controls and checks across customers without relying on intensive project work.