Seek.ai: Turning data questions into trustworthy answers

The customer meeting is tomorrow

A colleague needs data for a customer meeting the next day. Sarah Nagy stops her research, writes a straightforward data pull, and sends back the result. The colleague waits; the analyst loses time for deeper work. Repeated across her jobs, this interruption motivated [Seek.ai](https://www.seek.ai/): a natural-language interface that lets business users ask questions about company data.

In the interview published in March 2023, Nagy traces her career from research using Hubble Space Telescope data at UCLA and Caltech into quantitative finance. What excited her was discovering valuable patterns, including novel trading strategies. Routine requests consumed time she wanted to spend on those investigations.

Her experience with improv supplies a useful reliability analogy: an actor playing a doctor can sound convincing without saying anything medically true. Language models can likewise produce believable but false answers. For a data product, believability does not establish correctness—either in the answer or in the code used to obtain it.

Access to data does not teach judgment

Easier querying also makes it easier to find numbers that support an existing opinion. Nagy recalls a sell-side research analyst who tried different hyperparameter settings—the adjustable settings used in the analysis—and selected those that validated a qualitative thesis about a stock’s key performance indicator. The desired conclusion determined which analysis survived.

Nagy argues that being data-driven means using evidence to discover what is true and improve decisions. Giving people access is only part of the job; company leadership must also teach them how to interpret and use the insights they receive.

She positions Seek as automating tedious tasks that consume analysts’ time, with analysts continuing to do more valuable work. The target is a familiar automation gap: an ad hoc task feels repetitive enough to deserve a Python script, but attempts to automate it prove unexpectedly complicated. The analyst keeps doing it manually. Nagy expects this unwanted, time-consuming work to be an early foothold for enterprise AI.

From plausible queries to reusable business meaning

Turning a natural-language request into code introduces two distinct failures: - The generated code may fail to run. - It may run successfully and return incorrect data. Nagy describes collaboration between business users and the data team, including guardrails around generated code, as part of Seek’s response. The specific checks are not detailed.

This makes data automation substantially harder than wrapping a code-generation API in an application. Having data in a warehouse does not make useful answers easy to extract. The calculations and business meaning still need to be organized.

A semantic layer maps business metrics expressed in natural language to the SQL or Python code that calculates them. These mappings make calculations reusable: existing metrics become building blocks for additional ones. dbt is valued here as a place to organize the SQL behind business metrics. More detailed, mature semantic layers could then make those metrics easier for business users to access through natural language.

Reuse also applies to questions. When someone in a business has already asked a question, the interface should make it easy to retrieve that question and see its data. This is the concrete goal described in response to a host’s question about industry-specific suggestions: recover useful work already done within the business.

Freeing analysts for original investigation

The choice between dashboards and conversation depends on the work being done. Routine ad hoc requests are attractive candidates for automation when they require little new thinking and consume time better spent on research. Dashboard design, longer investigations, and deeper analysis can require original ideas—the work this account reserves an important human role for.

The argument is that predictive generation, trained on human-created material, is less suited to producing new ideas. That is a forecast about AI’s limitations, not an established boundary on creativity. The extension of this argument to image generation is explicitly tentative.

The practical aim is to free time for consequential investigation. The conversation returns to the excitement around big data circa 2015: writing code to analyze more information than a person could inspect manually, then finding insights that could improve financial decisions, increase revenue, or materially reduce costs.

Multiple models under an accuracy constraint

Seek describes a proprietary architecture containing multiple models, combining third-party and in-house software. External APIs and models adapted from GitHub or Hugging Face resources are possible components; customer preferences for particular vendors also influence selection. Specific models and their roles are not disclosed.

Its stated measure of performance is customer happiness: building something people want to use and buy, then improving it through customer conversations. The emphasis is on customer feedback and user experience, rather than solely on benchmark performance.

Asked whether users would wait longer for better answers, Nagy emphasizes accuracy. In previous roles, she gave business colleagues self-service tools to reduce manual requests, but found that people unintentionally misused them. Without enough training to notice the mistake, a user could base a consequential decision on bad data.

Her requirement is accurate answers without cutting corners. This is a product commitment; the interview provides neither a measured accuracy rate nor an acceptable response-time threshold.

The application code around the model call

Calling a model is only part of building a language-model application. Nagy describes an early completion-model failure: asked for a sentence about the American Revolution, the model might produce a paragraph, generate a new instruction about America, and answer that too. It continues the prompt-and-answer pattern beyond the requested response.

Early Seek development required handling code to cut output at the newline before that unwanted continuation. This explains LangChain’s appeal to Nagy: reusable components could save developers from repeatedly writing the machinery around model calls. She endorses that direction without claiming LangChain solves this particular case.

Customers also want analytics about Seek itself: how much it improves productivity and what return on investment it provides. Nagy says Seek tracks these kinds of metrics and could expose them through its own interface, though she gives no numerical results. The build-versus-buy principle is to concentrate internal effort on the customer-facing natural-language product and use existing infrastructure elsewhere when constraints allow.

That product vision includes conversation, with an exchange between the user and the system extending beyond a single search-bar submission. Nagy describes this as an important component already underway, without specifying its conversational mechanics.

A spaceship office—and everyday automation

In the lightning round, Nagy names Stable Diffusion as a favorite AI product. A simple prompt describing an AI spaceship office produced a Zoom background she liked, complete with a gaming chair, without elaborate rendering keywords. She also highlights [Eleuther AI](https://www.eleuther.ai/), whose Discord she had followed mostly as a lurker and whose apparent grassroots origins appealed to her, alongside LangChain and GPT Index.

Her prediction for the year following the interview is broad: rapid development and surprising projects, without claiming to know which ones will emerge. Her personal request is specific—better meeting scheduling, calendar management, and email automation. A host recommends Ipso.ai for coordinating multiple people’s calendars; Nagy says she will check it out.

Nagy closes with optimism about automating mundane work, including tasks people may not yet realize can be automated. She acknowledges potential harms, expresses confidence in efforts to address them, and encourages listeners to learn about both benefits and drawbacks. The opportunity she cares about is giving people back time for activities they find more rewarding.

At the close, she directs listeners to Seek.ai and identifies her Twitter handle as @SarahRNagy, with LinkedIn as another place to find her.